LayerCAM: Exploring Hierarchical Class Activation Maps for Localization

Know your orientation: A viewpoint-aware framework for polyp segmentation

Automatic polyp segmentation in endoscopic images is critical for the early diagnosis of colorectal cancer. Despite the availability of powerful segmentation models, two challenges still impede the accuracy of polyp segmentation algorithms. Firstly, during a colonoscopy, physicians frequently adjust the orientation of the colonoscope tip to capture underlying lesions, resulting in viewpoint changes in the colonoscopy images. These variations increase the diversity of polyp visual appearance, posing a challenge for learning robust polyp features. Secondly, polyps often exhibit properties similar to the surrounding tissues, leading to indistinct polyp boundaries. To address these problems, we propose a viewpoint-aware framework named VANet for precise polyp segmentation. In VANet, polyps are emphasized as a discriminative feature and thus can be localized by class activation maps in a viewpoint classification process. With these polyp locations, we design a viewpoint-aware Transformer (VAFormer) to alleviate the erosion of attention by the surrounding tissues, thereby inducing better polyp representations. Additionally, to enhance the polyp boundary perception of the network, we develop a boundary-aware Transformer (BAFormer) to encourage self-attention towards uncertain regions. As a consequence, the combination of the two modules is capable of calibrating predictions and significantly improving polyp segmentation performance. Extensive experiments on seven public datasets across six metrics demonstrate the state-of-the-art results of our method, and VANet can handle colonoscopy images in real-world scenarios effectively. The source code is available at https://github.com/1024803482/Viewpoint-Aware-Network.

Computer Vision Identification of Trachomatous Inflammation-Follicular Using Deep Learning

PURPOSE

Trachoma surveys are used to estimate the prevalence of trachomatous inflammation-follicular (TF) to guide mass antibiotic distribution. These surveys currently rely on human graders, introducing a significant resource burden and potential for human error. This study describes the development and evaluation of machine learning models intended to reduce cost and improve reliability of these surveys.

METHODS

Fifty-six thousand seven hundred twenty-five everted eyelid photographs were obtained from 11,358 children of age 0 to 9 years in a single trachoma-endemic region of Ethiopia over a 3-year period. Expert graders reviewed all images from each examination to determine the estimated number of tarsal conjunctival follicles and the degree of trachomatous inflammation-intense. The median estimate of the 3 grader groups was used as the ground truth to train a MobileNetV3 large deep convolutional neural network to detect cases with TF.

RESULTS

The classification model predicted a TF prevalence of 32%, which was not significantly different from the human consensus estimate (30%; 95% confidence interval of difference, -2 to +4%). The model had an area under the receiver operating characteristic curve of 0.943, F1 score of 0.923, 88% accuracy, 83% sensitivity, and 91% specificity. The area under the receiver operating characteristic curve increased to 0.995 when interpreting nonborderline cases of TF.

CONCLUSIONS

Deep convolutional neural network models performed well at classifying TF and detecting the number of follicles evident in conjunctival photographs. Implementation of similar models may enable accurate, efficient, large-scale trachoma screening. Further validation in diverse populations with varying TF prevalence is needed before implementation at scale.

Efficient diagnostic classification of diverse pathologies through contextual eye movement data analysis with a novel hybrid architecture

The analysis of eye movements has proven valuable for understanding brain function and the neuropathology of various disorders. This research aims to utilize eye movement data analysis as a screening tool for differentiation between eight different groups of pathologies, including scholar, neurologic, and postural disorders. Leveraging a dataset from 20 clinical centers, all employing AIDEAL and REMOBI eye movement technologies this study extends prior research by considering a multi-annotation setting, incorporating information from recordings from saccade and vergence eye movement tests, and using contextual information (e.g. target signals and latency of the eye movement relative to the target and confidence level of the quality of eye movement recording) to improve accuracy while reducing noise interference. Additionally, we introduce a novel hybrid architecture that combines the weight-sharing feature of convolution layers with the long-range capabilities of the transformer architecture to improve model efficiency and reduce the computation cost by a factor of 3.36, while still being competitive in terms of macro F1 score. Evaluated on two diverse datasets, our method demonstrates promising results, the most powerful discrimination being Attention & Neurologic; with a macro F1 score of up to 78.8%; disorder. The results indicate the effectiveness of our approach in classifying eye movement data from different pathologies and different clinical centers accurately, thus enabling the creation of an assistant tool in the future.

Modeling Chickpea Productivity with Artificial Image Objects and Convolutional Neural Network

The chickpea plays a significant role in global agriculture and occupies an increasing share in the human diet. The main aim of the research was to develop a model for the prediction of two chickpea productivity traits in the available dataset. Genomic data for accessions were encoded in Artificial Image Objects, and a model for the thousand-seed weight (TSW) and number of seeds per plant (SNpP) prediction was constructed using a Convolutional Neural Network, dictionary learning and sparse coding for feature extraction, and extreme gradient boosting for regression. The model was capable of predicting both traits with an acceptable accuracy of 84-85%. The most important factors for model solution were identified using the dense regression attention maps method. The SNPs important for the SNpP and TSW traits were found in 34 and 49 genes, respectively. Genomic prediction with a constructed model can help breeding programs harness genotypic and phenotypic diversity to more effectively produce varieties with a desired phenotype.

Gradient-Based Instance-Specific Visual Explanations for Object Specification and Object Discrimination

We propose the gradient-weighted Object Detector Activation Maps (ODAM), a visual explanation technique for interpreting the predictions of object detectors. Utilizing the gradients of detector targets flowing into the intermediate feature maps, ODAM produces heat maps that show the influence of regions on the detector's decision for each predicted attribute. Compared to previous works on classification activation maps (CAM), ODAM generates instance-specific explanations rather than class-specific ones. We show that ODAM is applicable to one-stage, two-stage, and transformer-based detectors with different types of detector backbones and heads, and produces higher-quality visual explanations than the state-of-the-art in terms of both effectiveness and efficiency. We discuss two explanation tasks for object detection: 1) object specification: what is the important region for the prediction? 2) object discrimination: which object is detected? Aiming at these two aspects, we present a detailed analysis of the visual explanations of detectors and carry out extensive experiments to validate the effectiveness of the proposed ODAM. Furthermore, we investigate user trust on the explanation maps, how well the visual explanations of object detectors agrees with human explanations, as measured through human eye gaze, and whether this agreement is related with user trust. Finally, we also propose two applications, ODAM-KD and ODAM-NMS, based on these two abilities of ODAM. ODAM-KD utilizes the object specification of ODAM to generate top-down attention for key predictions and instruct the knowledge distillation of object detection. ODAM-NMS considers the location of the model's explanation for each prediction to distinguish the duplicate detected objects. A training scheme, ODAM-Train, is proposed to improve the quality on object discrimination, and help with ODAM-NMS.

Deep Learning-Based Joint Effusion Classification in Adult Knee Radiographs: A Multi-Center Prospective Study

Knee effusion, a common and important indicator of joint diseases such as osteoarthritis, is typically more discernible on magnetic resonance imaging (MRI) scans compared to radiographs. However, the use of radiographs for the early detection of knee effusion remains promising due to their cost-effectiveness and accessibility. This multi-center prospective study collected a total of 1413 radiographs from four hospitals between February 2022 to March 2023, of which 1281 were analyzed after exclusions. To automatically detect knee effusion on radiographs, we utilized a state-of-the-art (SOTA) deep learning-based classification model with a novel preprocessing technique to optimize images for diagnosing knee effusion. The diagnostic performance of the proposed method was significantly higher than that of the baseline model, achieving an area under the receiver operating characteristic curve (AUC) of 0.892, accuracy of 0.803, sensitivity of 0.820, and specificity of 0.785. Moreover, the proposed method significantly outperformed two non-orthopedic physicians. Coupled with an explainable artificial intelligence method for visualization, this approach not only improved diagnostic performance but also interpretability, highlighting areas of effusion. These results demonstrate that the proposed method enables the early and accurate classification of knee effusions on radiographs, thereby reducing healthcare costs and improving patient outcomes through timely interventions.

The dynamic-static dual-branch deep neural network for urban speeding hotspot identification using street view image data

The visual information regarding the road environment can influence drivers' perception and judgment, often resulting in frequent speeding incidents. Identifying speeding hotspots in cities can prevent potential speeding incidents, thereby improving traffic safety levels. We propose the Dual-Branch Contextual Dynamic-Static Feature Fusion Network based on static panoramic images and dynamically changing sequence data, aiming to capture global features in the macro scene of the area and dynamically changing information in the micro view for a more accurate urban speeding hotspot area identification. For the static branch, we propose the Multi-scale Contextual Feature Aggregation Network for learning global spatial contextual association information. In the dynamic branch, we construct the Multi-view Dynamic Feature Fusion Network to capture the dynamically changing features of a scene from a continuous sequence of street view images. Additionally, we designed the Dynamic-Static Feature Correlation Fusion Structure to correlate and fuse dynamic and static features. The experimental results show that the model has good performance, and the overall recognition accuracy reaches 99.4%. The ablation experiments show that the recognition effect after the fusion of dynamic and static features is better than that of static and dynamic branches. The proposed model also shows better performance than other deep learning models. In addition, we combine image processing methods and different Class Activation Mapping (CAM) methods to extract speeding frequency visual features from the model perception results. The results show that more accurate speeding frequency features can be obtained by using LayerCAM and GradCAM-Plus for static global scenes and dynamic local sequences, respectively. In the static global scene, the speeding frequency features are mainly concentrated on the buildings and green layout on both sides of the road, while in the dynamic scene, the speeding frequency features shift with the scene changes and are mainly concentrated on the dynamically changing transition areas of greenery, roads, and surrounding buildings. The code and model used for identifying hotspots of urban traffic accidents in this study are available for access: https://github.com/gwt-ZJU/DCDSFF-Net.

Any region can be perceived equally and effectively on rotation pretext task using full rotation and weighted-region mixture

In recent years, self-supervised learning has emerged as a powerful approach to learning visual representations without requiring extensive manual annotation. One popular technique involves using rotation transformations of images, which provide a clear visual signal for learning semantic representation. However, in this work, we revisit the pretext task of predicting image rotation in self-supervised learning and discover that it tends to marginalise the perception of features located near the centre of an image. To address this limitation, we propose a new self-supervised learning method, namely FullRot, which spotlights underrated regions by resizing the randomly selected and cropped regions of images. Moreover, FullRot increases the complexity of the rotation pretext task by applying the degree-free rotation to the region cropped into a circle. To encourage models to learn from different general parts of an image, we introduce a new data mixture technique called WRMix, which merges two random intra-image patches. By combining these innovative crop and rotation methods with the data mixture scheme, our approach, FullRot + WRMix, surpasses the state-of-the-art self-supervision methods in classification, segmentation, and object detection tasks on ten benchmark datasets with an improvement of up to +13.98% accuracy on STL-10, +8.56% accuracy on CIFAR-10, +10.20% accuracy on Sports-100, +15.86% accuracy on Mammals-45, +15.15% accuracy on PAD-UFES-20, +32.44% mIoU on VOC 2012, +7.62% mIoU on ISIC 2018, +9.70% mIoU on FloodArea, +25.16% AP50 on VOC 2007, and +58.69% AP50 on UTDAC 2020. The code is available at https://github.com/anthonyweidai/FullRot_WRMix.

Clinical domain knowledge-derived template improves post hoc AI explanations in pneumothorax classification

OBJECTIVE

Pneumothorax is an acute thoracic disease caused by abnormal air collection between the lungs and chest wall. Recently, artificial intelligence (AI), especially deep learning (DL), has been increasingly employed for automating the diagnostic process of pneumothorax. To address the opaqueness often associated with DL models, explainable artificial intelligence (XAI) methods have been introduced to outline regions related to pneumothorax. However, these explanations sometimes diverge from actual lesion areas, highlighting the need for further improvement.

METHOD

We propose a template-guided approach to incorporate the clinical knowledge of pneumothorax into model explanations generated by XAI methods, thereby enhancing the quality of the explanations. Utilizing one lesion delineation created by radiologists, our approach first generates a template that represents potential areas of pneumothorax occurrence. This template is then superimposed on model explanations to filter out extraneous explanations that fall outside the template's boundaries. To validate its efficacy, we carried out a comparative analysis of three XAI methods (Saliency Map, Grad-CAM, and Integrated Gradients) with and without our template guidance when explaining two DL models (VGG-19 and ResNet-50) in two real-world datasets (SIIM-ACR and ChestX-Det).

RESULTS

The proposed approach consistently improved baseline XAI methods across twelve benchmark scenarios built on three XAI methods, two DL models, and two datasets. The average incremental percentages, calculated by the performance improvements over the baseline performance, were 97.8% in Intersection over Union (IoU) and 94.1% in Dice Similarity Coefficient (DSC) when comparing model explanations and ground-truth lesion areas. We further visualized baseline and template-guided model explanations on radiographs to showcase the performance of our approach.

CONCLUSIONS

In the context of pneumothorax diagnoses, we proposed a template-guided approach for improving model explanations. Our approach not only aligns model explanations more closely with clinical insights but also exhibits extensibility to other thoracic diseases. We anticipate that our template guidance will forge a novel approach to elucidating AI models by integrating clinical domain expertise.

Overlapping Shoeprint Detection by Edge Detection and Deep Learning

In the field of 2-D image processing and computer vision, accurately detecting and segmenting objects in scenarios where they overlap or are obscured remains a challenge. This difficulty is worse in the analysis of shoeprints used in forensic investigations because they are embedded in noisy environments such as the ground and can be indistinct. Traditional convolutional neural networks (CNNs), despite their success in various image analysis tasks, struggle with accurately delineating overlapping objects due to the complexity of segmenting intertwined textures and boundaries against a background of noise. This study introduces and employs the YOLO (You Only Look Once) model enhanced by edge detection and image segmentation techniques to improve the detection of overlapping shoeprints. By focusing on the critical boundary information between shoeprint textures and the ground, our method demonstrates improvements in sensitivity and precision, achieving confidence levels above 85% for minimally overlapped images and maintaining above 70% for extensively overlapped instances. Heatmaps of convolution layers were generated to show how the network converges towards successful detection using these enhancements. This research may provide a potential methodology for addressing the broader challenge of detecting multiple overlapping objects against noisy backgrounds.

Deep learning to detect left ventricular structural abnormalities in chest X-rays

BACKGROUND AND AIMS

Early identification of cardiac structural abnormalities indicative of heart failure is crucial to improving patient outcomes. Chest X-rays (CXRs) are routinely conducted on a broad population of patients, presenting an opportunity to build scalable screening tools for structural abnormalities indicative of Stage B or worse heart failure with deep learning methods. In this study, a model was developed to identify severe left ventricular hypertrophy (SLVH) and dilated left ventricle (DLV) using CXRs.

METHODS

A total of 71 589 unique CXRs from 24 689 different patients completed within 1 year of echocardiograms were identified. Labels for SLVH, DLV, and a composite label indicating the presence of either were extracted from echocardiograms. A deep learning model was developed and evaluated using area under the receiver operating characteristic curve (AUROC). Performance was additionally validated on 8003 CXRs from an external site and compared against visual assessment by 15 board-certified radiologists.

RESULTS

The model yielded an AUROC of 0.79 (0.76-0.81) for SLVH, 0.80 (0.77-0.84) for DLV, and 0.80 (0.78-0.83) for the composite label, with similar performance on an external data set. The model outperformed all 15 individual radiologists for predicting the composite label and achieved a sensitivity of 71% vs. 66% against the consensus vote across all radiologists at a fixed specificity of 73%.

CONCLUSIONS

Deep learning analysis of CXRs can accurately detect the presence of certain structural abnormalities and may be useful in early identification of patients with LV hypertrophy and dilation. As a resource to promote further innovation, 71 589 CXRs with adjoining echocardiographic labels have been made publicly available.

XRelevanceCAM: towards explainable tissue characterization with improved localisation of pathological structures in probe-based confocal laser endomicroscopy

PURPOSE

Probe-based confocal laser endomicroscopy (pCLE) enables intraoperative tissue characterization with improved resection rates of brain tumours. Although a plethora of deep learning models have been developed for automating tissue characterization, their lack of transparency is a concern. To tackle this issue, techniques like Class Activation Map (CAM) and its variations highlight image regions related to model decisions. However, they often fall short of providing human-interpretable visual explanations for surgical decision support, primarily due to the shattered gradient problem or insufficient theoretical underpinning.

METHODS

In this paper, we introduce XRelevanceCAM, an explanation method rooted in a better backpropagation approach, incorporating sensitivity and conservation axioms. This enhanced method offers greater theoretical foundation and effectively mitigates the shattered gradient issue when compared to other CAM variants.

RESULTS

Qualitative and quantitative evaluations are based on ex vivo pCLE data of brain tumours. XRelevanceCAM effectively highlights clinically relevant areas that characterize the tissue type. Specifically, it yields a remarkable 56% improvement over our closest baseline, RelevanceCAM, in the network's shallowest layer as measured by the mean Intersection over Union (mIoU) metric based on ground-truth annotations (from 18 to 28.07%). Furthermore, a 6% improvement in mIoU is observed when generating the final saliency map from all network layers.

CONCLUSION

We introduce a new CAM variation, XRelevanceCAM, for precise identification of clinically important structures in pCLE data. This can aid introperative decision support in brain tumour resection surgery, as validated in our performance study.

Advancing musculoskeletal tumor diagnosis: Automated segmentation and predictive classification using deep learning and radiomics

OBJECTIVES

Musculoskeletal (MSK) tumors, given their high mortality rate and heterogeneity, necessitate precise examination and diagnosis to guide clinical treatment effectively. Magnetic resonance imaging (MRI) is pivotal in detecting MSK tumors, as it offers exceptional image contrast between bone and soft tissue. This study aims to enhance the speed of detection and the diagnostic accuracy of MSK tumors through automated segmentation and grading utilizing MRI.

MATERIALS AND METHODS

The research included 170 patients (mean age, 58 years ±12 (standard deviation), 84 men) with MSK lesions, who underwent MRI scans from April 2021 to May 2023. We proposed a deep learning (DL) segmentation model MSAPN based on multi-scale attention and pixel-level reconstruction, and compared it with existing algorithms. Using MSAPN-segmented lesions to extract their radiomic features for the benign and malignant classification of tumors.

RESULTS

Compared to the most advanced segmentation algorithms, MSAPN demonstrates better performance. The Dice similarity coefficients (DSC) are 0.871 and 0.815 in the testing set and independent validation set, respectively. The radiomics model for classifying benign and malignant lesions achieves an accuracy of 0.890. Moreover, there is no statistically significant difference between the radiomics model based on manual segmentation and MSAPN segmentation.

CONCLUSION

This research contributes to the advancement of MSK tumor diagnosis through automated segmentation and predictive classification. The integration of DL algorithms and radiomics shows promising results, and the visualization analysis of feature maps enhances clinical interpretability.

Better Understanding Differences in Attribution Methods via Systematic Evaluations

Deep neural networks are very successful on many vision tasks, but hard to interpret due to their black box nature. To overcome this, various post-hoc attribution methods have been proposed to identify image regions most influential to the models' decisions. Evaluating such methods is challenging since no ground truth attributions exist. We thus propose three novel evaluation schemes to more reliably measure the faithfulness of those methods, to make comparisons between them more fair, and to make visual inspection more systematic. To address faithfulness, we propose a novel evaluation setting (DiFull) in which we carefully control which parts of the input can influence the output in order to distinguish possible from impossible attributions. To address fairness, we note that different methods are applied at different layers, which skews any comparison, and so evaluate all methods on the same layers (ML-Att) and discuss how this impacts their performance on quantitative metrics. For more systematic visualizations, we propose a scheme (AggAtt) to qualitatively evaluate the methods on complete datasets. We use these evaluation schemes to study strengths and shortcomings of some widely used attribution methods over a wide range of models. Finally, we propose a post-processing smoothing step that significantly improves the performance of some attribution methods, and discuss its applicability.

Machine learning with multimodal neuroimaging data to classify stages of Alzheimer's disease: a systematic review and meta-analysis

In recent years, Alzheimer's disease (AD) has been a serious threat to human health. Researchers and clinicians alike encounter a significant obstacle when trying to accurately identify and classify AD stages. Several studies have shown that multimodal neuroimaging input can assist in providing valuable insights into the structural and functional changes in the brain related to AD. Machine learning (ML) algorithms can accurately categorize AD phases by identifying patterns and linkages in multimodal neuroimaging data using powerful computational methods. This study aims to assess the contribution of ML methods to the accurate classification of the stages of AD using multimodal neuroimaging data. A systematic search is carried out in IEEE Xplore, Science Direct/Elsevier, ACM DigitalLibrary, and PubMed databases with forward snowballing performed on Google Scholar. The quantitative analysis used 47 studies. The explainable analysis was performed on the classification algorithm and fusion methods used in the selected studies. The pooled sensitivity and specificity, including diagnostic efficiency, were evaluated by conducting a meta-analysis based on a bivariate model with the hierarchical summary receiver operating characteristics (ROC) curve of multimodal neuroimaging data and ML methods in the classification of AD stages. Wilcoxon signed-rank test is further used to statistically compare the accuracy scores of the existing models. With a 95% confidence interval of 78.87-87.71%, the combined sensitivity for separating participants with mild cognitive impairment (MCI) from healthy control (NC) participants was 83.77%; for separating participants with AD from NC, it was 94.60% (90.76%, 96.89%); for separating participants with progressive MCI (pMCI) from stable MCI (sMCI), it was 80.41% (74.73%, 85.06%). With a 95% confidence interval (78.87%, 87.71%), the Pooled sensitivity for distinguishing mild cognitive impairment (MCI) from healthy control (NC) participants was 83.77%, with a 95% confidence interval (90.76%, 96.89%), the Pooled sensitivity for distinguishing AD from NC was 94.60%, likewise (MCI) from healthy control (NC) participants was 83.77% progressive MCI (pMCI) from stable MCI (sMCI) was 80.41% (74.73%, 85.06%), and early MCI (EMCI) from NC was 86.63% (82.43%, 89.95%). Pooled specificity for differentiating MCI from NC was 79.16% (70.97%, 87.71%), AD from NC was 93.49% (91.60%, 94.90%), pMCI from sMCI was 81.44% (76.32%, 85.66%), and EMCI from NC was 85.68% (81.62%, 88.96%). The Wilcoxon signed rank test showed a low P-value across all the classification tasks. Multimodal neuroimaging data with ML is a promising future in classifying the stages of AD but more research is required to increase the validity of its application in clinical practice.

Explainability of three-dimensional convolutional neural networks for functional magnetic resonance imaging of Alzheimer's disease classification based on gradient-weighted class activation mapping

Currently, numerous studies focus on employing fMRI-based deep neural networks to diagnose neurological disorders such as Alzheimer's Disease (AD), yet only a handful have provided results regarding explainability. We address this gap by applying several prevalent explainability methods such as gradient-weighted class activation mapping (Grad-CAM) to an fMRI-based 3D-VGG16 network for AD diagnosis to improve the model's explainability. The aim is to explore the specific Region of Interest (ROI) of brain the model primarily focuses on when making predictions, as well as whether there are differences in these ROIs between AD and normal controls (NCs). First, we utilized multiple resting-state functional activity maps including ALFF, fALFF, ReHo, and VMHC to reduce the complexity of fMRI data, which differed from many studies that utilized raw fMRI data. Compared to methods utilizing raw fMRI data, this manual feature extraction approach may potentially alleviate the model's burden. Subsequently, 3D-VGG16 were employed for AD classification, where the final fully connected layers were replaced with a Global Average Pooling (GAP) layer, aimed at mitigating overfitting while preserving spatial information within the feature maps. The model achieved a maximum of 96.4% accuracy on the test set. Finally, several 3D CAM methods were employed to interpret the models. In the explainability results of the models with relatively high accuracy, the highlighted ROIs were primarily located in the precuneus and the hippocampus for AD subjects, while the models focused on the entire brain for NC. This supports current research on ROIs involved in AD. We believe that explaining deep learning models would not only provide support for existing research on brain disorders, but also offer important referential recommendations for the study of currently unknown etiologies.

Magnetoencephalography Decoding Transfer Approach: From Deep Learning Models to Intrinsically Interpretable Models

When decoding neuroelectrophysiological signals represented by Magnetoencephalography (MEG), deep learning models generally achieve high predictive performance but lack the ability to interpret their predicted results. This limitation prevents them from meeting the essential requirements of reliability and ethical-legal considerations in practical applications. In contrast, intrinsically interpretable models, such as decision trees, possess self-evident interpretability while typically sacrificing accuracy. To effectively combine the respective advantages of both deep learning and intrinsically interpretable models, an MEG transfer approach through feature attribution-based knowledge distillation is pioneered, which transforms deep models (teacher) into highly accurate intrinsically interpretable models (student). The resulting models provide not only intrinsic interpretability but also high predictive performance, besides serving as an excellent approximate proxy to understand the inner workings of deep models. In the proposed approach, post-hoc feature knowledge derived from post-hoc interpretable algorithms, specifically feature attribution maps, is introduced into knowledge distillation for the first time. By guiding intrinsically interpretable models to assimilate this knowledge, the transfer of MEG decoding information from deep models to intrinsically interpretable models is implemented. Experimental results demonstrate that the proposed approach outperforms the benchmark knowledge distillation algorithms. This approach successfully improves the prediction accuracy of Soft Decision Tree by a maximum of 8.28%, reaching almost equivalent or even superior performance to deep teacher models. Furthermore, the model-agnostic nature of this approach offers broad application potential.

R-Cut: Enhancing Explainability in Vision Transformers with Relationship Weighted Out and Cut

Transformer-based models have gained popularity in the field of natural language processing (NLP) and are extensively utilized in computer vision tasks and multi-modal models such as GPT4. This paper presents a novel method to enhance the explainability of transformer-based image classification models. Our method aims to improve trust in classification results and empower users to gain a deeper understanding of the model for downstream tasks by providing visualizations of class-specific maps. We introduce two modules: the "Relationship Weighted Out" and the "Cut" modules. The "Relationship Weighted Out" module focuses on extracting class-specific information from intermediate layers, enabling us to highlight relevant features. Additionally, the "Cut" module performs fine-grained feature decomposition, taking into account factors such as position, texture, and color. By integrating these modules, we generate dense class-specific visual explainability maps. We validate our method with extensive qualitative and quantitative experiments on the ImageNet dataset. Furthermore, we conduct a large number of experiments on the LRN dataset, which is specifically designed for automatic driving danger alerts, to evaluate the explainability of our method in scenarios with complex backgrounds. The results demonstrate a significant improvement over previous methods. Moreover, we conduct ablation experiments to validate the effectiveness of each module. Through these experiments, we are able to confirm the respective contributions of each module, thus solidifying the overall effectiveness of our proposed approach.

Enhancing Retinal Fundus Image Quality Assessment With Swin-Transformer-Based Learning Across Multiple Color-Spaces

Purpose

The assessment of retinal image (RI) quality holds significant importance in both clinical trials and large datasets, because suboptimal images can potentially conceal early signs of diseases, thereby resulting in inaccurate medical diagnoses. This study aims to develop an automatic method for Retinal Image Quality Assessment (RIQA) that incorporates visual explanations, aiming to comprehensively evaluate the quality of retinal fundus images (RIs).

Methods

We developed an automatic RIQA system, named Swin-MCSFNet, utilizing 28,792 RIs from the EyeQ dataset, as well as 2000 images from the EyePACS dataset and an additional 1,000 images from the OIA-ODIR dataset. After preprocessing, including cropping black regions, data augmentation, and normalization, a Swin-MCSFNet classifier based on the Swin-Transformer for multiple color-space fusion was proposed to grade the quality of RIs. The generalizability of Swin-MCSFNet was validated across multiple data centers. Additionally, for enhanced interpretability, a Score-CAM-generated heatmap was applied to provide visual explanations.

Results

Experimental results reveal that the proposed Swin-MCSFNet achieves promising performance, yielding a micro-receiver operating characteristic (ROC) of 0.93 and ROC scores of 0.96, 0.81, and 0.96 for the "Good," "Usable," and "Reject" categories, respectively. These scores underscore the accuracy of RIQA based on Swin-MCSF in distinguishing among the three categories. Furthermore, heatmaps generated across different RIQA classification scores and various color spaces suggest that regions in the retinal images from multiple color spaces contribute significantly to the decision-making process of the Swin-MCSFNet classifier.

Conclusions

Our study demonstrates that the proposed Swin-MCSFNet outperforms other methods in experiments conducted on multiple datasets, as evidenced by the superior performance metrics and insightful Score-CAM heatmaps.

Translational Relevance

This study constructs a new retinal image quality evaluation system, which will contribute to the subsequent research of retinal images.

Convolutional neural network model for automatic recognition and classification of pancreatic cancer cell based on analysis of lipid droplet on unlabeled sample by 3D optical diffraction tomography

INTRODUCTION

Pancreatic cancer cells generally accumulate large numbers of lipid droplets (LDs), which regulate lipid storage. To promote rapid diagnosis, an automatic pancreatic cancer cell recognition system based on a deep convolutional neural network was proposed in this study using quantitative images of LDs from stain-free cytologic samples by optical diffraction tomography.

METHODS

We retrieved 3D refractive index tomograms and reconstructed 37 optical images of one cell. From the four cell lines, the obtained fields were separated into training and test datasets with 10,397 and 3,478 images, respectively. Furthermore, we adopted several machine learning techniques based on a single image-based prediction model to improve the performance of the computer-aided diagnostic system.

RESULTS

Pancreatic cancer cells had a significantly lower total cell volume and dry mass than did normal pancreatic cells and were accompanied by greater numbers of lipid droplets (LDs). When evaluating multitask learning techniques utilizing the EfficientNet-b3 model through confusion matrices, the overall 2-category accuracy for cancer classification reached 96.7 %. Simultaneously, the overall 4-category accuracy for individual cell line classification achieved a high accuracy of 96.2 %. Furthermore, when we added the core techniques one by one, the overall performance of the proposed technique significantly improved, reaching an area under the curve (AUC) of 0.997 and an accuracy of 97.06 %. Finally, the AUC reached 0.998 through the ablation study with the score fusion technique.

DISCUSSION

Our novel training strategy has significant potential for automating and promoting rapid recognition of pancreatic cancer cells. In the near future, deep learning-embedded medical devices will substitute laborious manual cytopathologic examinations for sustainable economic potential.

Evidence-based XAI: An empirical approach to design more effective and explainable decision support systems

This paper proposes a user study aimed at evaluating the impact of Class Activation Maps (CAMs) as an eXplainable AI (XAI) method in a radiological diagnostic task, the detection of thoracolumbar (TL) fractures from vertebral X-rays. In particular, we focus on two oft-neglected features of CAMs, that is granularity and coloring, in terms of what features, lower-level vs higher-level, should the maps highlight and adopting which coloring scheme, to bring better impact to the decision-making process, both in terms of diagnostic accuracy (that is effectiveness) and of user-centered dimensions, such as perceived confidence and utility (that is satisfaction), depending on case complexity, AI accuracy, and user expertise. Our findings show that lower-level features CAMs, which highlight more focused anatomical landmarks, are associated with higher diagnostic accuracy than higher-level features CAMs, particularly among experienced physicians. Moreover, despite the intuitive appeal of semantic CAMs, traditionally colored CAMs consistently yielded higher diagnostic accuracy across all groups. Our results challenge some prevalent assumptions in the XAI field and emphasize the importance of adopting an evidence-based and human-centered approach to design and evaluate AI- and XAI-assisted diagnostic tools. To this aim, the paper also proposes a hierarchy of evidence framework to help designers and practitioners choose the XAI solutions that optimize performance and satisfaction on the basis of the strongest evidence available or to focus on the gaps in the literature that need to be filled to move from opinionated and eminence-based research to one more based on empirical evidence and end-user work and preferences.

Stratification of tumour cell radiation response and metabolic signatures visualization with Raman spectroscopy and explainable convolutional neural network

Reprogramming of cellular metabolism is a driving factor of tumour progression and radiation therapy resistance. Identifying biochemical signatures associated with tumour radioresistance may assist with the development of targeted treatment strategies to improve clinical outcomes. Raman spectroscopy (RS) can monitor post-irradiation biomolecular changes and signatures of radiation response in tumour cells in a label-free manner. Convolutional Neural Networks (CNN) perform feature extraction directly from data in an end-to-end learning manner, with high classification performance. Furthermore, recently developed CNN explainability techniques help visualize the critical discriminative features captured by the model. In this work, a CNN is developed to characterize tumour response to radiotherapy based on its degree of radioresistance. The model was trained to classify Raman spectra of three human tumour cell lines as radiosensitive (LNCaP) or radioresistant (MCF7, H460) over a range of treatment doses and data collection time points. Additionally, a method based on Gradient-Weighted Class Activation Mapping (Grad-CAM) was used to determine response-specific salient Raman peaks influencing the CNN predictions. The CNN effectively classified the cell spectra, with accuracy, sensitivity, specificity, and F1 score exceeding 99.8%. Grad-CAM heatmaps of H460 and MCF7 cell spectra (radioresistant) exhibited high contributions from Raman bands tentatively assigned to glycogen, amino acids, and nucleic acids. Conversely, heatmaps of LNCaP cells (radiosensitive) revealed activations at lipid and phospholipid bands. Finally, Grad-CAM variable importance scores were derived for glycogen, asparagine, and phosphatidylcholine, and we show that their trends over cell line, dose, and acquisition time agreed with previously established models. Thus, the CNN can accurately detect biomolecular differences in the Raman spectra of tumour cells of varying radiosensitivity without requiring manual feature extraction. Finally, Grad-CAM may help identify metabolic signatures associated with the observed categories, offering the potential for automated clinical tumour radiation response characterization.

DeepGraFT: A novel semantic segmentation auxiliary ROI-based deep learning framework for effective fundus tessellation classification

Fundus tessellation (FT) is a prevalent clinical feature associated with myopia and has implications in the development of myopic maculopathy, which causes irreversible visual impairment. Accurate classification of FT in color fundus photo can help predict the disease progression and prognosis. However, the lack of precise detection and classification tools has created an unmet medical need, underscoring the importance of exploring the clinical utility of FT. Thus, to address this gap, we introduce an automatic FT grading system (called DeepGraFT) using classification-and-segmentation co-decision models by deep learning. ConvNeXt, utilizing transfer learning from pretrained ImageNet weights, was employed for the classification algorithm, aligning with a region of interest based on the ETDRS grading system to boost performance. A segmentation model was developed to detect FT exits, complementing the classification for improved grading accuracy. The training set of DeepGraFT was from our in-house cohort (MAGIC), and the validation sets consisted of the rest part of in-house cohort and an independent public cohort (UK Biobank). DeepGraFT demonstrated a high performance in the training stage and achieved an impressive accuracy in validation phase (in-house cohort: 86.85 %; public cohort: 81.50 %). Furthermore, our findings demonstrated that DeepGraFT surpasses machine learning-based classification models in FT classification, achieving a 5.57 % increase in accuracy. Ablation analysis revealed that the introduced modules significantly enhanced classification effectiveness and elevated accuracy from 79.85 % to 86.85 %. Further analysis using the results provided by DeepGraFT unveiled a significant negative association between FT and spherical equivalent (SE) in the UK Biobank cohort. In conclusion, DeepGraFT accentuates potential benefits of the deep learning model in automating the grading of FT and allows for potential utility as a clinical-decision support tool for predicting progression of pathological myopia.

Weakly Supervised Solar Panel Mapping via Uncertainty Adjusted Label Transition in Aerial Images

This paper proposes a novel uncertainty-adjusted label transition (UALT) method for weakly supervised solar panel mapping (WS-SPM) in aerial Images. In weakly supervised learning (WSL), the noisy nature of pseudo labels (PLs) often leads to poor model performance. To address this problem, we formulate the task as a label-noise learning problem and build a statistically consistent mapping model by estimating the instance-dependent transition matrix (IDTM). We propose to estimate the IDTM with a parameterized label transition network describing the relationship between the latent clean labels and noisy PLs. A trace regularizer is employed to impose constraints on the form of IDTM for its stability. To further reduce the estimation difficulty of IDTM, we incorporate uncertainty estimation to first improve the accuracy of noisy dataset distillation and then mitigate the negative impacts of falsely distilled examples with an uncertainty-adjusted re-weighting strategy. Extensive experiments and ablation studies on two challenging aerial data sets support the validity of the proposed UALT.

Enhancing Nigrosome-1 Sign Identification via Interpretable AI using True Susceptibility Weighted Imaging

BACKGROUND

Nigrosome 1 (N1), the largest nigrosome region in the ventrolateral area of the substantia nigra pars compacta, is identifiable by the "N1 sign" in long echo time gradient echo MRI. The N1 sign's absence is a vital Parkinson's disease (PD) diagnostic marker. However, it is challenging to visualize and assess the N1 sign in clinical practice.

PURPOSE

To automatically detect the presence or absence of the N1 sign from true susceptibility weighted imaging by using deep-learning method.

STUDY TYPE

Prospective.

POPULATION/SUBJECTS

453 subjects, including 225 PD patients, 120 healthy controls (HCs), and 108 patients with other movement disorders, were prospectively recruited including 227 males and 226 females. They were divided into training, validation, and test cohorts of 289, 73, and 91 cases, respectively.

FIELD STRENGTH/SEQUENCE

3D gradient echo SWI sequence at 3T; 3D multiecho strategically acquired gradient echo imaging at 3T; NM-sensitive 3D gradient echo sequence with MTC pulse at 3T.

ASSESSMENT

A neuroradiologist with 5 years of experience manually delineated substantia nigra regions. Two raters with 2 and 36 years of experience assessed the N1 sign on true susceptibility weighted imaging (tSWI), QSM with high-pass filter, and magnitude data combined with MTC data. We proposed NINet, a neural model, for automatic N1 sign identification in tSWI images.

STATISTICAL TESTS

We compared the performance of NINet to the subjective reference standard using Receiver Operating Characteristic analyses, and a decision curve analysis assessed identification accuracy.

RESULTS

NINet achieved an area under the curve (AUC) of 0.87 (CI: 0.76-0.89) in N1 sign identification, surpassing other models and neuroradiologists. NINet localized the putative N1 sign within tSWI images with 67.3% accuracy.

DATA CONCLUSION

Our proposed NINet model's capability to determine the presence or absence of the N1 sign, along with its localization, holds promise for enhancing diagnostic accuracy when evaluating PD using MR images.

LEVEL OF EVIDENCE

2 TECHNICAL EFFICACY: Stage 1.

Convolutional Neural Networks for Glioma Segmentation and Prognosis: A Systematic Review

Deep learning (DL) is poised to redefine the way medical images are processed and analyzed. Convolutional neural networks (CNNs), a specific type of DL architecture, are exceptional for high-throughput processing, allowing for the effective extraction of relevant diagnostic patterns from large volumes of complex visual data. This technology has garnered substantial interest in the field of neuro-oncology as a promising tool to enhance medical imaging throughput and analysis. A multitude of methods harnessing MRI-based CNNs have been proposed for brain tumor segmentation, classification, and prognosis prediction. They are often applied to gliomas, the most common primary brain cancer, to classify subtypes with the goal of guiding therapy decisions. Additionally, the difficulty of repeating brain biopsies to evaluate treatment response in the setting of often confusing imaging findings provides a unique niche for CNNs to help distinguish the treatment response to gliomas. For example, glioblastoma, the most aggressive type of brain cancer, can grow due to poor treatment response, can appear to grow acutely due to treatment-related inflammation as the tumor dies (pseudo-progression), or falsely appear to be regrowing after treatment as a result of brain damage from radiation (radiation necrosis). CNNs are being applied to separate this diagnostic dilemma. This review provides a detailed synthesis of recent DL methods and applications for intratumor segmentation, glioma classification, and prognosis prediction. Furthermore, this review discusses the future direction of MRI-based CNN in the field of neuro-oncology and challenges in model interpretability, data availability, and computation efficiency.

STAMarker: determining spatial domain-specific variable genes with saliency maps in deep learning

Spatial transcriptomics characterizes gene expression profiles while retaining the information of the spatial context, providing an unprecedented opportunity to understand cellular systems. One of the essential tasks in such data analysis is to determine spatially variable genes (SVGs), which demonstrate spatial expression patterns. Existing methods only consider genes individually and fail to model the inter-dependence of genes. To this end, we present an analytic tool STAMarker for robustly determining spatial domain-specific SVGs with saliency maps in deep learning. STAMarker is a three-stage ensemble framework consisting of graph-attention autoencoders, multilayer perceptron (MLP) classifiers, and saliency map computation by the backpropagated gradient. We illustrate the effectiveness of STAMarker and compare it with serveral commonly used competing methods on various spatial transcriptomic data generated by different platforms. STAMarker considers all genes at once and is more robust when the dataset is very sparse. STAMarker could identify spatial domain-specific SVGs for characterizing spatial domains and enable in-depth analysis of the region of interest in the tissue section.

An empirical comparison of deep learning explainability approaches for EEG using simulated ground truth

Recent advancements in machine learning and deep learning (DL) based neural decoders have significantly improved decoding capabilities using scalp electroencephalography (EEG). However, the interpretability of DL models remains an under-explored area. In this study, we compared multiple model explanation methods to identify the most suitable method for EEG and understand when some of these approaches might fail. A simulation framework was developed to evaluate the robustness and sensitivity of twelve back-propagation-based visualization methods by comparing to ground truth features. Multiple methods tested here showed reliability issues after randomizing either model weights or labels: e.g., the saliency approach, which is the most used visualization technique in EEG, was not class or model-specific. We found that DeepLift was consistently accurate as well as robust to detect the three key attributes tested here (temporal, spatial, and spectral precision). Overall, this study provides a review of model explanation methods for DL-based neural decoders and recommendations to understand when some of these methods fail and what they can capture in EEG.

Ground truth based comparison of saliency maps algorithms

Deep neural networks (DNNs) have achieved outstanding results in domains such as image processing, computer vision, natural language processing and bioinformatics. In recent years, many methods have been proposed that can provide a visual explanation of decision made by such classifiers. Saliency maps are probably the most popular. However, it is still unclear how to properly interpret saliency maps for a given image and which techniques perform most accurately. This paper presents a methodology to practically evaluate the real effectiveness of saliency map generation methods. We used three state-of-the-art network architectures along with specially prepared benchmark datasets, and we proposed a novel metric to provide a quantitative comparison of the methods. The comparison identified the most reliable techniques and the solutions which usually failed in our tests.

UniVisNet: A Unified Visualization and Classification Network for accurate grading of gliomas from MRI

Accurate grading of brain tumors plays a crucial role in the diagnosis and treatment of glioma. While convolutional neural networks (CNNs) have shown promising performance in this task, their clinical applicability is still constrained by the interpretability and robustness of the models. In the conventional framework, the classification model is trained first, and then visual explanations are generated. However, this approach often leads to models that prioritize classification performance or complexity, making it difficult to achieve a precise visual explanation. Motivated by these challenges, we propose the Unified Visualization and Classification Network (UniVisNet), a novel framework that aims to improve both the classification performance and the generation of high-resolution visual explanations. UniVisNet addresses attention misalignment by introducing a subregion-based attention mechanism, which replaces traditional down-sampling operations. Additionally, multiscale feature maps are fused to achieve higher resolution, enabling the generation of detailed visual explanations. To streamline the process, we introduce the Unified Visualization and Classification head (UniVisHead), which directly generates visual explanations without the need for additional separation steps. Through extensive experiments, our proposed UniVisNet consistently outperforms strong baseline classification models and prevalent visualization methods. Notably, UniVisNet achieves remarkable results on the glioma grading task, including an AUC of 94.7%, an accuracy of 89.3%, a sensitivity of 90.4%, and a specificity of 85.3%. Moreover, UniVisNet provides visually interpretable explanations that surpass existing approaches. In conclusion, UniVisNet innovatively generates visual explanations in brain tumor grading by simultaneously improving the classification performance and generating high-resolution visual explanations. This work contributes to the clinical application of deep learning, empowering clinicians with comprehensive insights into the spatial heterogeneity of glioma.

AI vs. AI: Can AI Detect AI-Generated Images?

The proliferation of Artificial Intelligence (AI) models such as Generative Adversarial Networks (GANs) has shown impressive success in image synthesis. Artificial GAN-based synthesized images have been widely spread over the Internet with the advancement in generating naturalistic and photo-realistic images. This might have the ability to improve content and media; however, it also constitutes a threat with regard to legitimacy, authenticity, and security. Moreover, implementing an automated system that is able to detect and recognize GAN-generated images is significant for image synthesis models as an evaluation tool, regardless of the input modality. To this end, we propose a framework for reliably detecting AI-generated images from real ones through Convolutional Neural Networks (CNNs). First, GAN-generated images were collected based on different tasks and different architectures to help with the generalization. Then, transfer learning was applied. Finally, several Class Activation Maps (CAM) were integrated to determine the discriminative regions that guided the classification model in its decision. Our approach achieved 100% on our dataset, i.e., Real or Synthetic Images (RSI), and a superior performance on other datasets and configurations in terms of its accuracy. Hence, it can be used as an evaluation tool in image generation. Our best detector was a pre-trained EfficientNetB4 fine-tuned on our dataset with a batch size of 64 and an initial learning rate of 0.001 for 20 epochs. Adam was used as an optimizer, and learning rate reduction along with data augmentation were incorporated.

Automatic segmentation of neurovascular bundle on mri using deep learning based topological modulated network

PURPOSE

Radiation damage on neurovascular bundles (NVBs) may be the cause of sexual dysfunction after radiotherapy for prostate cancer. However, it is challenging to delineate NVBs as organ-at-risks from planning CTs during radiotherapy. Recently, the integration of MR into radiotherapy made NVBs contour delineating possible. In this study, we aim to develop an MRI-based deep learning method for automatic NVB segmentation.

METHODS

The proposed method, named topological modulated network, consists of three subnetworks, that is, a focal modulation, a hierarchical block and a topological fully convolutional network (FCN). The focal modulation is used to derive the location and bounds of left and right NVBs', namely the candidate volume-of-interests (VOIs). The hierarchical block aims to highlight the NVB boundaries information on derived feature map. The topological FCN then segments the NVBs inside the VOIs by considering the topological consistency nature of the vascular delineating. Based on the location information of candidate VOIs, the segmentations of NVBs can then be brought back to the input MRI's coordinate system.

RESULTS

A five-fold cross-validation study was performed on 60 patient cases to evaluate the performance of the proposed method. The segmented results were compared with manual contours. The Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95 ) are (left NVB) 0.81 ± 0.10, 1.49 ± 0.88 mm, and (right NVB) 0.80 ± 0.15, 1.54 ± 1.22 mm, respectively.

CONCLUSION

We proposed a novel deep learning-based segmentation method for NVBs on pelvic MR images. The good segmentation agreement of our method with the manually drawn ground truth contours supports the feasibility of the proposed method, which can be potentially used to spare NVBs during proton and photon radiotherapy and thereby improve the quality of life for prostate cancer patients.

Weakly Supervised Breast Lesion Detection in Dynamic Contrast-Enhanced MRI

Currently, obtaining accurate medical annotations requires high labor and time effort, which largely limits the development of supervised learning-based tumor detection tasks. In this work, we investigated a weakly supervised learning model for detecting breast lesions in dynamic contrast-enhanced MRI (DCE-MRI) with only image-level labels. Two hundred fifty-four normal and 398 abnormal cases with pathologically confirmed lesions were retrospectively enrolled into the breast dataset, which was divided into the training set (80%), validation set (10%), and testing set (10%) at the patient level. First, the second image series S2 after the injection of a contrast agent was acquired from the 3.0-T, T1-weighted dynamic enhanced MR imaging sequences. Second, a feature pyramid network (FPN) with convolutional block attention module (CBAM) was proposed to extract multi-scale feature maps of the modified classification network VGG16. Then, initial location information was obtained from the heatmaps generated using the layer class activation mapping algorithm (Layer-CAM). Finally, the detection results of breast lesion were refined by the conditional random field (CRF). Accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC) were utilized for evaluation of image-level classification. Average precision (AP) was estimated for breast lesion localization. Delong's test was used to compare the AUCs of different models for significance. The proposed model was effective with accuracy of 95.2%, sensitivity of 91.6%, specificity of 99.2%, and AUC of 0.986. The AP for breast lesion detection was 84.1% using weakly supervised learning. Weakly supervised learning based on FPN combined with Layer-CAM facilitated automatic detection of breast lesion.

CSF-Glioma: A Causal Segmentation Framework for Accurate Grading and Subregion Identification of Gliomas

Deep networks have shown strong performance in glioma grading; however, interpreting their decisions remains challenging due to glioma heterogeneity. To address these challenges, the proposed solution is the Causal Segmentation Framework (CSF). This framework aims to accurately predict high- and low-grade gliomas while simultaneously highlighting key subregions. Our framework utilizes a shrinkage segmentation method to identify subregions containing essential decision information. Moreover, we introduce a glioma grading module that combines deep learning and traditional approaches for precise grading. Our proposed model achieves the best performance among all models, with an AUC of 96.14%, an F1 score of 93.74%, an accuracy of 91.04%, a sensitivity of 91.83%, and a specificity of 88.88%. Additionally, our model exhibits efficient resource utilization, completing predictions within 2.31s and occupying only 0.12 GB of memory during the test phase. Furthermore, our approach provides clear and specific visualizations of key subregions, surpassing other methods in terms of interpretability. In conclusion, the Causal Segmentation Framework (CSF) demonstrates its effectiveness at accurately predicting glioma grades and identifying key subregions. The inclusion of causality in the CSF model enhances the reliability and accuracy of preoperative decision-making for gliomas. The interpretable results provided by the CSF model can assist clinicians in their assessment and treatment planning.

DPDH-CapNet: A Novel Lightweight Capsule Network with Non-routing for COVID-19 Diagnosis Using X-ray Images

COVID-19 has claimed millions of lives since its outbreak in December 2019, and the damage continues, so it is urgent to develop new technologies to aid its diagnosis. However, the state-of-the-art deep learning methods often rely on large-scale labeled data, limiting their clinical application in COVID-19 identification. Recently, capsule networks have achieved highly competitive performance for COVID-19 detection, but they require expensive routing computation or traditional matrix multiplication to deal with the capsule dimensional entanglement. A more lightweight capsule network is developed to effectively address these problems, namely DPDH-CapNet, which aims to enhance the technology of automated diagnosis for COVID-19 chest X-ray images. It adopts depthwise convolution (D), point convolution (P), and dilated convolution (D) to construct a new feature extractor, thus successfully capturing the local and global dependencies of COVID-19 pathological features. Simultaneously, it constructs the classification layer by homogeneous (H) vector capsules with an adaptive, non-iterative, and non-routing mechanism. We conduct experiments on two publicly available combined datasets, including normal, pneumonia, and COVID-19 images. With a limited number of samples, the parameters of the proposed model are reduced by 9x compared to the state-of-the-art capsule network. Moreover, our model has faster convergence speed and better generalization, and its accuracy, precision, recall, and F-measure are improved to 97.99%, 98.05%, 98.02%, and 98.03%, respectively. In addition, experimental results demonstrate that, contrary to the transfer learning method, the proposed model does not require pre-training and a large number of training samples.

Hybrid Domain Consistency Constraints-Based Deep Neural Network for Facial Expression Recognition

Facial expression recognition (FER) has received increasing attention. However, multiple factors (e.g., uneven illumination, facial deflection, occlusion, and subjectivity of annotations in image datasets) probably reduce the performance of traditional FER methods. Thus, we propose a novel Hybrid Domain Consistency Network (HDCNet) based on a feature constraint method that combines both spatial domain consistency and channel domain consistency. Specifically, first, the proposed HDCNet mines the potential attention consistency feature expression (different from manual features, e.g., HOG and SIFT) as effective supervision information by comparing the original sample image with the augmented facial expression image. Second, HDCNet extracts facial expression-related features in the spatial and channel domains, and then it constrains the consistent expression of features through the mixed domain consistency loss function. In addition, the loss function based on the attention-consistency constraints does not require additional labels. Third, the network weights are learned to optimize the classification network through the loss function of the mixed domain consistency constraints. Finally, experiments conducted on the public RAF-DB and AffectNet benchmark datasets verify that the proposed HDCNet improved classification accuracy by 0.3-3.84% compared to the existing methods.

DGA3-Net: A parameter-efficient deep learning model for ASPECTS assessment for acute ischemic stroke using non-contrast computed tomography

Detecting the early signs of stroke using non-contrast computerized tomography (NCCT) is essential for the diagnosis of acute ischemic stroke (AIS). However, the hypoattenuation in NCCT is difficult to precisely identify, and accurate assessments of the Alberta Stroke Program Early CT Score (ASPECTS) are usually time-consuming and require experienced neuroradiologists. To this end, this study proposes DGA3-Net, a convolutional neural network (CNN)-based model for ASPECTS assessment via detecting early ischemic changes in ASPECTS regions. DGA3-Net is based on a novel parameter-efficient dihedral group CNN encoder to exploit the rotation and reflection symmetry of convolution kernels. The bounding volume of each ASPECTS region is extracted from the encoded feature, and an attention-guided slice aggregation module is used to aggregate features from all slices. An asymmetry-aware classifier is then used to predict stroke presence via comparison between ASPECTS regions from the left and right hemispheres. Pre-treatment NCCTs of suspected AIS patients were collected retrospectively, which consists of a primary dataset (n = 170) and an external validation dataset (n = 90), with expert consensus ASPECTS readings as ground truth. DGA3-Net outperformed two expert neuroradiologists in regional stroke identification (F1 = 0.69) and ASPECTS evaluation (Cohen's weighted Kappa = 0.70). Our ablation study also validated the efficacy of the proposed model design. In addition, class-relevant areas highlighted by visualization techniques corresponded highly with various well-established qualitative imaging signs, further validating the learned representation. This study demonstrates the potential of deep learning techniques for timely and accurate AIS diagnosis from NCCT, which could substantially improve the quality of treatment for AIS patients.

Multimodal deep neural decoding reveals highly resolved spatiotemporal profile of visual object representation in humans

Perception and categorization of objects in a visual scene are essential to grasp the surrounding situation. Recently, neural decoding schemes, such as machine learning in functional magnetic resonance imaging (fMRI), has been employed to elucidate the underlying neural mechanisms. However, it remains unclear as to how spatially distributed brain regions temporally represent visual object categories and sub-categories. One promising strategy to address this issue is neural decoding with concurrently obtained neural response data of high spatial and temporal resolution. In this study, we explored the spatial and temporal organization of visual object representations using concurrent fMRI and electroencephalography (EEG), combined with neural decoding using deep neural networks (DNNs). We hypothesized that neural decoding by multimodal neural data with DNN would show high classification performance in visual object categorization (faces or non-face objects) and sub-categorization within faces and objects. Visualization of the fMRI DNN was more sensitive than that in the univariate approach and revealed that visual categorization occurred in brain-wide regions. Interestingly, the EEG DNN valued the earlier phase of neural responses for categorization and the later phase of neural responses for sub-categorization. Combination of the two DNNs improved the classification performance for both categorization and sub-categorization compared with fMRI DNN or EEG DNN alone. These deep learning-based results demonstrate a categorization principle in which visual objects are represented in a spatially organized and coarse-to-fine manner, and provide strong evidence of the ability of multimodal deep learning to uncover spatiotemporal neural machinery in sensory processing.

Classification and localization of maize leaf spot disease based on weakly supervised learning

Precisely discerning disease types and vulnerable areas is crucial in implementing effective monitoring of crop production. This forms the basis for generating targeted plant protection recommendations and automatic, precise applications. In this study, we constructed a dataset comprising six types of field maize leaf images and developed a framework for classifying and localizing maize leaf diseases. Our approach involved integrating lightweight convolutional neural networks with interpretable AI algorithms, which resulted in high classification accuracy and fast detection speeds. To evaluate the performance of our framework, we tested the mean Intersection over Union (mIoU) of localized disease spot coverage and actual disease spot coverage when relying solely on image-level annotations. The results showed that our framework achieved a mIoU of up to 55.302%, indicating the feasibility of using weakly supervised semantic segmentation based on class activation mapping techniques for identifying disease spots in crop disease detection. This approach, which combines deep learning models with visualization techniques, improves the interpretability of the deep learning models and achieves successful localization of infected areas of maize leaves through weakly supervised learning. The framework allows for smart monitoring of crop diseases and plant protection operations using mobile phones, smart farm machines, and other devices. Furthermore, it offers a reference for deep learning research on crop diseases.

Cross-modal guiding and reweighting network for multi-modal RSVP-based target detection

Rapid Serial Visual Presentation (RSVP) based Brain-Computer Interface (BCI) facilities the high-throughput detection of rare target images by detecting evoked event-related potentials (ERPs). At present, the decoding accuracy of the RSVP-based BCI system limits its practical applications. This study introduces eye movements (gaze and pupil information), referred to as EYE modality, as another useful source of information to combine with EEG-based BCI and forms a novel target detection system to detect target images in RSVP tasks. We performed an RSVP experiment, recorded the EEG signals and eye movements simultaneously during a target detection task, and constructed a multi-modal dataset including 20 subjects. Also, we proposed a cross-modal guiding and fusion network to fully utilize EEG and EYE modalities and fuse them for better RSVP decoding performance. In this network, a two-branch backbone was built to extract features from these two modalities. A Cross-Modal Feature Guiding (CMFG) module was proposed to guide EYE modality features to complement the EEG modality for better feature extraction. A Multi-scale Multi-modal Reweighting (MMR) module was proposed to enhance the multi-modal features by exploring intra- and inter-modal interactions. And, a Dual Activation Fusion (DAF) was proposed to modulate the enhanced multi-modal features for effective fusion. Our proposed network achieved a balanced accuracy of 88.00% (±2.29) on the collected dataset. The ablation studies and visualizations revealed the effectiveness of the proposed modules. This work implies the effectiveness of introducing the EYE modality in RSVP tasks. And, our proposed network is a promising method for RSVP decoding and further improves the performance of RSVP-based target detection systems.

MSMFN: An Ultrasound Based Multi-Step Modality Fusion Network for Identifying the Histologic Subtypes of Metastatic Cervical Lymphadenopathy

Identifying squamous cell carcinoma and adenocarcinoma subtypes of metastatic cervical lymphadenopathy (CLA) is critical for localizing the primary lesion and initiating timely therapy. B-mode ultrasound (BUS), color Doppler flow imaging (CDFI), ultrasound elastography (UE) and dynamic contrast-enhanced ultrasound provide effective tools for identification but synthesis of modality information is a challenge for clinicians. Therefore, based on deep learning, rationally fusing these modalities with clinical information to personalize the classification of metastatic CLA requires new explorations. In this paper, we propose Multi-step Modality Fusion Network (MSMFN) for multi-modal ultrasound fusion to identify histological subtypes of metastatic CLA. MSMFN can mine the unique features of each modality and fuse them in a hierarchical three-step process. Specifically, first, under the guidance of high-level BUS semantic feature maps, information in CDFI and UE is extracted by modality interaction, and the static imaging feature vector is obtained. Then, a self-supervised feature orthogonalization loss is introduced to help learn modality heterogeneity features while maintaining maximal task-consistent category distinguishability of modalities. Finally, six encoded clinical information are utilized to avoid prediction bias and improve prediction ability further. Our three-fold cross-validation experiments demonstrate that our method surpasses clinicians and other multi-modal fusion methods with an accuracy of 80.06%, a true-positive rate of 81.81%, and a true-negative rate of 80.00%. Our network provides a multi-modal ultrasound fusion framework that considers prior clinical knowledge and modality-specific characteristics. Our code will be available at: https://github.com/RichardSunnyMeng/MSMFN.

Endoscopic Image Classification Based on Explainable Deep Learning

Deep learning has achieved remarkably positive results and impacts on medical diagnostics in recent years. Due to its use in several proposals, deep learning has reached sufficient accuracy to implement; however, the algorithms are black boxes that are hard to understand, and model decisions are often made without reason or explanation. To reduce this gap, explainable artificial intelligence (XAI) offers a huge opportunity to receive informed decision support from deep learning models and opens the black box of the method. We conducted an explainable deep learning method based on ResNet152 combined with Grad-CAM for endoscopy image classification. We used an open-source KVASIR dataset that consisted of a total of 8000 wireless capsule images. The heat map of the classification results and an efficient augmentation method achieved a high positive result with 98.28% training and 93.46% validation accuracy in terms of medical image classification.

Towards Query Efficient Black-Box Attacks: A Universal Dual Transferability-Based Framework

Adversarial attacks have threatened the application of deep neural networks in security-sensitive scenarios. Most existing black-box attacks fool the target model by interacting with it many times and producing global perturbations. However, all pixels are not equally crucial to the target model, so indiscriminately treating all pixels will increase the query overhead inevitably. Besides, existing black-box attacks take clean samples as start points, which also limits the query efficiency. In this paper, we propose a novel black-box attack framework, constructed on a strategy of dual transferability (DT), to perturb the discriminative areas of clean examples within limited queries. The first kind of transferability is the transferability of model interpretations. Based on this property, we identify the discriminative areas of clean samples for generating local perturbations. The second one is the transferability of adversarial examples, which helps us to produce local pre-perturbations for further improving the query efficiency. We achieve the two kinds of transferability through an independent auxiliary model and do not incur extra query overhead. After identifying discriminative areas and generating pre-perturbations, we use the pre-perturbed samples as better start points and further perturb them locally in a black-box way to search the corresponding adversarial examples. The DT strategy is general, so the proposed framework can be applied to different types of black-box attacks. We conduct extensive experiments to show that, under various system settings, our framework can significantly improve the query efficiency of existing black-box attacks and attack success rate.

Weakly supervised semantic segmentation for MRI: exploring the advantages and disadvantages of class activation maps for biological image segmentation with soft boundaries

Fully supervised semantic segmentation models require pixel-level annotations that are costly to obtain. As a remedy, weakly supervised semantic segmentation has been proposed, where image-level labels and class activation maps (CAM) can detect discriminative regions for specific class objects. In this paper, we evaluated several CAM methods applied to different convolutional neural networks (CNN) to highlight tissue damage of cod fillets with soft boundaries in MRI. Our results show that different CAM methods produce very different CAM regions, even when applying them to the same CNN model. CAM methods that claim to highlight more of the class object do not necessarily highlight more damaged regions or originate from the same high discriminatory regions, nor do these damaged regions show high agreement across the different CAM methods. Additionally, CAM methods produce damaged regions that do not align with external reference metrics, and even show correlations contrary to what can be expected.

TSSK-Net: Weakly supervised biomarker localization and segmentation with image-level annotation in retinal OCT images

The localization and segmentation of biomarkers in OCT images are critical steps in retina-related disease diagnosis. Although fully supervised deep learning models can segment pathological regions, their performance relies on labor-intensive pixel-level annotations. Compared with dense pixel-level annotation, image-level annotation can reduce the burden of manual annotation. Existing methods for image-level annotation are usually based on class activation maps (CAM). However, current methods still suffer from model collapse, training instability, and anatomical mismatch due to the considerable variation in retinal biomarkers' shape, texture, and size. This paper proposes a novel weakly supervised biomarkers localization and segmentation method, requiring only image-level annotations. The technique is a Teacher-Student network with joint Self-supervised contrastive learning and Knowledge distillation-based anomaly localization, namely TSSK-Net. Specifically, we treat retinal biomarker regions as abnormal regions distinct from normal regions. First, we propose a novel pre-training strategy based on supervised contrastive learning that encourages the model to learn the anatomical structure of normal OCT images. Second, we design a fine-tuning module and propose a novel hybrid network structure. The network includes supervised contrastive loss for feature learning and cross-entropy loss for classification learning. To further improve the performance, we propose an efficient strategy to combine these two losses to preserve the anatomical structure and enhance the encoding representation of features. Finally, we design a knowledge distillation-based anomaly segmentation method that is effectively combined with the previous model to alleviate the challenge of insufficient supervision. Experimental results on a local dataset and a public dataset demonstrated the effectiveness of our proposed method. Our proposed method can effectively reduce the annotation burden of ophthalmologists in OCT images.

Research on Assistant Diagnosis of Fundus Optic Neuropathy Based on Deep Learning

PURPOSE

The purpose of this study was to use the neural network to distinguish optic edema (ODE), and optic atrophy from normal fundus images and try to use visualization to explain the artificial intelligence methods.

METHODS

Three hundred and sixty-seven images of ODE, 206 images of optic atrophy, and 231 images of normal fundus were used, which were provided by two hospitals. A set of image preprocessing and data enhancement methods were created and a variety of different neural network models, such as VGG16, VGG19, Inception V3, and 50-layer Deep Residual Learning (ResNet50) were used. The accuracy, recall, F1-score, and ROC curve under different networks were analyzed to evaluate the performance of models. Besides, CAM (class activation mapping) was utilized to find the focus of neural network and visualization of neural network with feature fusion.

RESULTS

Our image preprocessing and data enhancement method significantly improved the accuracy of model performance by about 10%. Among the networks, VGG16 had the best effect, as the accuracy of ODE, optic atrophy and normal fundus were 98, 90, and 95%, respectively. The macro-average and micro-average of VGG16 both reached 0.98. From CAM we can clearly find out that the focus area of the network is near the optic cup. From feature fusion images, we can find out the difference between the three types fundus images.

CONCLUSION

Through image preprocessing, data enhancement, and neural network training, we applied artificial intelligence to identify ophthalmic diseases, acquired the focus area through CAM, and identified the difference between the three ophthalmic diseases through neural network middle layers visualization. With the help of assistant diagnosis, ophthalmologists can evaluate cases more precisely and more clearly.