Explainable AI to improve acceptance of convolutional neural networks for automatic classification of dopamine transporter SPECT in the diagnosis of clinically uncertain parkinsonian syndromes

Automated identification of uncertain cases in deep learning-based classification of dopamine transporter SPECT to improve clinical utility and acceptance

PURPOSE

Deep convolutional neural networks (CNN) are promising for automatic classification of dopamine transporter (DAT)-SPECT images. Reporting the certainty of CNN-based decisions is highly desired to flag cases that might be misclassified and, therefore, require particularly careful inspection by the user. The aim of the current study was to design and validate a CNN-based system for the identification of uncertain cases.

METHODS

A network ensemble (NE) combining five CNNs was trained for binary classification of [123I]FP-CIT DAT-SPECT images as "normal" or "neurodegeneration-typical reduction" with high accuracy (NE for classification, NEfC). An uncertainty detection module (UDM) was obtained by combining two additional NE, one trained for detection of "reduced" DAT-SPECT with high sensitivity, the other with high specificity. A case was considered "uncertain" if the "high sensitivity" NE and the "high specificity" NE disagreed. An internal "development" dataset of 1740 clinical DAT-SPECT images was used for training (n = 1250) and testing (n = 490). Two independent datasets with different image characteristics were used for testing only (n = 640, 645). Three established approaches for uncertainty detection were used for comparison (sigmoid, dropout, model averaging).

RESULTS

In the test data from the development dataset, the NEfC achieved 98.0% accuracy. 4.3% of all test cases were flagged as "uncertain" by the UDM: 2.5% of the correctly classified cases and 90% of the misclassified cases. NEfC accuracy among "certain" cases was 99.8%. The three comparison methods were less effective in labelling misclassified cases as "uncertain" (40-80%). These findings were confirmed in both additional test datasets.

CONCLUSION

The UDM allows reliable identification of uncertain [123I]FP-CIT SPECT with high risk of misclassification. We recommend that automatic classification of [123I]FP-CIT SPECT images is combined with an UDM to improve clinical utility and acceptance. The proposed UDM method ("high sensitivity versus high specificity") might be useful also for DAT imaging with other ligands and for other binary classification tasks.

A scoping review of interpretability and explainability concerning artificial intelligence methods in medical imaging

PURPOSE

To review eXplainable Artificial Intelligence/(XAI) methods available for medical imaging/(MI).

METHOD

A scoping review was conducted following the Joanna Briggs Institute's methodology. The search was performed on Pubmed, Embase, Cinhal, Web of Science, BioRxiv, MedRxiv, and Google Scholar. Studies published in French and English after 2017 were included. Keyword combinations and descriptors related to explainability, and MI modalities were employed. Two independent reviewers screened abstracts, titles and full text, resolving differences through discussion.

RESULTS

228 studies met the criteria. XAI publications are increasing, targeting MRI (n = 73), radiography (n = 47), CT (n = 46). Lung (n = 82) and brain (n = 74) pathologies, Covid-19 (n = 48), Alzheimer's disease (n = 25), brain tumors (n = 15) are the main pathologies explained. Explanations are presented visually (n = 186), numerically (n = 67), rule-based (n = 11), textually (n = 11), and example-based (n = 6). Commonly explained tasks include classification (n = 89), prediction (n = 47), diagnosis (n = 39), detection (n = 29), segmentation (n = 13), and image quality improvement (n = 6). The most frequently provided explanations were local (78.1 %), 5.7 % were global, and 16.2 % combined both local and global approaches. Post-hoc approaches were predominantly employed. The used terminology varied, sometimes indistinctively using explainable (n = 207), interpretable (n = 187), understandable (n = 112), transparent (n = 61), reliable (n = 31), and intelligible (n = 3).

CONCLUSION

The number of XAI publications in medical imaging is increasing, primarily focusing on applying XAI techniques to MRI, CT, and radiography for classifying and predicting lung and brain pathologies. Visual and numerical output formats are predominantly used. Terminology standardisation remains a challenge, as terms like "explainable" and "interpretable" are sometimes being used indistinctively. Future XAI development should consider user needs and perspectives.

Diagnosis of thyroid disease using deep convolutional neural network models applied to thyroid scintigraphy images: a multicenter study

Objectives

The aim of this study was to improve the diagnostic performance of nuclear medicine physicians using a deep convolutional neural network (DCNN) model and validate the results with two multicenter datasets for thyroid disease by analyzing clinical single-photon emission computed tomography (SPECT) image data.

Methods

In this multicenter retrospective study, 3194 SPECT thyroid images were collected for model training (n=2067), internal validation (n=514) and external validation (n=613). First, four pretrained DCNN models (AlexNet, ShuffleNetV2, MobileNetV3 and ResNet-34) for were tested multiple medical image classification of thyroid disease types (i.e., Graves' disease, subacute thyroiditis, thyroid tumor and normal thyroid). The best performing model was then subjected to fivefold cross-validation to further assess its performance, and the diagnostic performance of this model was compared with that of junior and senior nuclear medicine physicians. Finally, class-specific attentional regions were visualized with attention heatmaps using gradient-weighted class activation mapping.

Results

Each of the four pretrained neural networks attained an overall accuracy of more than 0.85 for the classification of SPECT thyroid images. The improved ResNet-34 model performed best, with an accuracy of 0.944. For the internal validation set, the ResNet-34 model showed higher accuracy (p < 0.001) when compared to that of the senior nuclear medicine physician, with an improvement of nearly 10%. Our model achieved an overall accuracy of 0.931 for the external dataset, a significantly higher accuracy than that of the senior physician (0.931 vs. 0.868, p < 0.001).

Conclusion

The DCNN-based model performed well in terms of diagnosing thyroid scintillation images. The DCNN model showed higher sensitivity and greater specificity in identifying Graves' disease, subacute thyroiditis, and thyroid tumors compared to those of nuclear medicine physicians, illustrating the feasibility of deep learning models to improve the diagnostic efficiency for assisting clinicians.

Survey of explainable artificial intelligence techniques for biomedical imaging with deep neural networks

Artificial Intelligence (AI) techniques of deep learning have revolutionized the disease diagnosis with their outstanding image classification performance. In spite of the outstanding results, the widespread adoption of these techniques in clinical practice is still taking place at a moderate pace. One of the major hindrance is that a trained Deep Neural Networks (DNN) model provides a prediction, but questions about why and how that prediction was made remain unanswered. This linkage is of utmost importance for the regulated healthcare domain to increase the trust in the automated diagnosis system by the practitioners, patients and other stakeholders. The application of deep learning for medical imaging has to be interpreted with caution due to the health and safety concerns similar to blame attribution in the case of an accident involving autonomous cars. The consequences of both a false positive and false negative cases are far reaching for patients' welfare and cannot be ignored. This is exacerbated by the fact that the state-of-the-art deep learning algorithms comprise of complex interconnected structures, millions of parameters, and a 'black box' nature, offering little understanding of their inner working unlike the traditional machine learning algorithms. Explainable AI (XAI) techniques help to understand model predictions which help develop trust in the system, accelerate the disease diagnosis, and meet adherence to regulatory requirements. This survey provides a comprehensive review of the promising field of XAI for biomedical imaging diagnostics. We also provide a categorization of the XAI techniques, discuss the open challenges, and provide future directions for XAI which would be of interest to clinicians, regulators and model developers.

Stereotactical normalization with multiple templates representative of normal and Parkinson-typical reduction of striatal uptake improves the discriminative power of automatic semi-quantitative analysis in dopamine transporter SPECT

BACKGROUND

The specific binding ratio (SBR) of ¹²³I-FP-CIT in the putamen is widely used to support the interpretation of dopamine transporter (DAT) SPECT. Automatic methods for computation of the putamen SBR often include stereotactical normalization of the individual DAT-SPECT image to an anatomical standard space. This study compared using a single ¹²³I-FP-CIT template image as target for stereotactical normalization versus multiple templates representative of normal and different levels of Parkinson-typical reduction of striatal ¹²³I-FP-CIT uptake.

METHODS

1702 clinical ¹²³I-FP-CIT SPECT images were stereotactically normalized (affine) to the anatomical space of the Montreal Neurological Institute (MNI) with SPM12 either using a single custom-made ¹²³I-FP-CIT template representative of normal striatal uptake or using eight different templates representative of normal and different levels of Parkinson-typical reduction of striatal FP-CIT uptake with and without attenuation and scatter correction. In the latter case, SPM finds the linear combination of the multiple templates that best matches the patient's image. The putamen SBR was obtained using hottest voxels analysis in large unilateral regions-of-interest predefined in MNI space. The histogram of the putamen SBR in the whole sample was fitted by the sum of two Gaussians. The power to differentiate between reduced and normal SBR was estimated by the effect size of the distance between the two Gaussians computed as the differences between their mean values scaled to their pooled standard deviation.

RESULTS

The effect size of the distance between the two Gaussians was 3.83 with the single template versus 3.96 with multiple templates for stereotactical normalization.

CONCLUSIONS

Multiple templates representative of normal and different levels of Parkinson-typical reduction for stereotactical normalization of DAT-SPECT might provide improved separation between normal and reduced putamen SBR that could result in slightly improved power for the detection of nigrostriatal degeneration.

Real-World Testing of a Machine Learning-Derived Visual Scale for Tc99m TRODAT-1 for Diagnosing Lewy Body Disease: Comparison with a Traditional Approach Using Semiquantification

Objectives: Abnormal dopamine transporter (DAT) uptake is an important biomarker for diagnosing Lewy body disease (LBD), including Parkinson’s disease (PD) and dementia with Lewy bodies (DLB). We evaluated a machine learning-derived visual scale (ML-VS) for Tc99m TRODAT-1 from one center and compared it with the striatal/background ratio (SBR) using semiquantification for diagnosing LBD in two other centers. Patients and Methods: This was a retrospective analysis of data from a history-based computerized dementia diagnostic system. MT-VS and SBR among normal controls (NCs) and patients with PD, PD with dementia (PDD), DLB, or Alzheimer’s disease (AD) were compared. Results: We included 715 individuals, including 122 NCs, 286 patients with PD, 40 with AD, 179 with DLB, and 88 with PDD. Compared with NCs, patients with PD exhibited a significantly higher prevalence of abnormal DAT uptake using all methods. Compared with the AD group, PDD and DLB groups exhibited a significantly higher prevalence of abnormal DAT uptake using all methods. The distribution of ML-VS was significantly different between PD and NC, DLB and AD, and PDD and AD groups (all p < 0.001). The correlation coefficient of ML-VS/SBR in all participants was 0.679. Conclusions: The ML-VS designed in one center is useful for differentiating PD from NC, DLB from AD, and PDD from AD in other centers. Its correlation with traditional approaches using different scanning machines is also acceptable. Future studies should develop models using data pools from multiple centers for increasing diagnostic accuracy.

An Explainable Classification Method of SPECT Myocardial Perfusion Images in Nuclear Cardiology Using Deep Learning and Grad-CAM

Background: This study targets the development of an explainable deep learning methodology for the automatic classification of coronary artery disease, utilizing SPECT MPI images. Deep learning is currently judged as non-transparent due to the model’s complex non-linear structure, and thus, it is considered a «black box», making it hard to gain a comprehensive understanding of its internal processes and explain its behavior. Existing explainable artificial intelligence tools can provide insights into the internal functionality of deep learning and especially of convolutional neural networks, allowing transparency and interpretation. Methods: This study seeks to address the identification of patients’ CAD status (infarction, ischemia or normal) by developing an explainable deep learning pipeline in the form of a handcrafted convolutional neural network. The proposed RGB-CNN model utilizes various pre- and post-processing tools and deploys a state-of-the-art explainability tool to produce more interpretable predictions in decision making. The dataset includes cases from 625 patients as stress and rest representations, comprising 127 infarction, 241 ischemic, and 257 normal cases previously classified by a doctor. The imaging dataset was split into 20% for testing and 80% for training, of which 15% was further used for validation purposes. Data augmentation was employed to increase generalization. The efficacy of the well-known Grad-CAM-based color visualization approach was also evaluated in this research to provide predictions with interpretability in the detection of infarction and ischemia in SPECT MPI images, counterbalancing any lack of rationale in the results extracted by the CNNs. Results: The proposed model achieved 93.3% accuracy and 94.58% AUC, demonstrating efficient performance and stability. Grad-CAM has shown to be a valuable tool for explaining CNN-based judgments in SPECT MPI images, allowing nuclear physicians to make fast and confident judgments by using the visual explanations offered. Conclusions: Prediction results indicate a robust and efficient model based on the deep learning methodology which is proposed for CAD diagnosis in nuclear medicine.

Soft Attention Based DenseNet Model for Parkinson’s Disease Classification Using SPECT Images

Objective

Deep learning algorithms have long been involved in the diagnosis of severe neurological disorders that interfere with patients’ everyday tasks, such as Parkinson’s disease (PD). The most effective imaging modality for detecting the condition is DaTscan, a variety of single-photon emission computerized tomography (SPECT) imaging method. The goal is to create a convolutional neural network that can specifically identify the region of interest following feature extraction.

Methods

The study comprised a total of 1,390 DaTscan imaging groups with PD and normal classes. The architecture of DenseNet-121 is leveraged with a soft-attention block added before the final classification layer. For visually analyzing the region of interest (ROI) from the images after classification, Soft Attention Maps and feature map representation are used.

Outcomes

The model obtains an overall accuracy of 99.2% and AUC-ROC score 99%. A sensitivity of 99.2%, specificity of 99.4% and f1-score of 99.1% is achieved that surpasses all prior research findings. Soft-attention map and feature map representation aid in highlighting the ROI, with a specific attention on the putamen and caudate regions.

Conclusion

With the deep learning framework adopted, DaTscan images reveal the putamen and caudate areas of the brain, which aid in the distinguishing of normal and PD cohorts with high accuracy and sensitivity.

Deep Learning-Based Automated Diagnosis for Coronary Artery Disease Using SPECT-MPI Images

(1) Background: Single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) is a long-established estimation methodology for medical diagnosis using image classification illustrating conditions in coronary artery disease. For these procedures, convolutional neural networks have proven to be very beneficial in achieving near-optimal accuracy for the automatic classification of SPECT images. (2) Methods: This research addresses the supervised learning-based ideal observer image classification utilizing an RGB-CNN model in heart images to diagnose CAD. For comparison purposes, we employ VGG-16 and DenseNet-121 pre-trained networks that are indulged in an image dataset representing stress and rest mode heart states acquired by SPECT. In experimentally evaluating the method, we explore a wide repertoire of deep learning network setups in conjunction with various robust evaluation and exploitation metrics. Additionally, to overcome the image dataset cardinality restrictions, we take advantage of the data augmentation technique expanding the set into an adequate number. Further evaluation of the model was performed via 10-fold cross-validation to ensure our model's reliability. (3) Results: The proposed RGB-CNN model achieved an accuracy of 91.86%, while VGG-16 and DenseNet-121 reached 88.54% and 86.11%, respectively. (4) Conclusions: The abovementioned experiments verify that the newly developed deep learning models may be of great assistance in nuclear medicine and clinical decision-making.

Predicting the Risk of Dental Implant Loss Using Deep Learning

AIM

To investigate the feasibility of predicting dental implant loss risk with deep learning (DL) based on preoperative cone-beam computed tomography.

MATERIALS AND METHODS

Six hundred and three patients who underwent implant surgery (279 high-risk patients who did and 324 low-risk patients who did not experience implant loss within 5 years) from January 2012 to January 2020 were enrolled. Three models, a logistic regression clinical model (CM) based on clinical features, a DL model based on radiography features, and an integrated model (IM) developed by combining CM with DL, were developed to predict the 5-year implant loss risk. The area under the receiver operating characteristic curve (AUC) was used to evaluate the model performance. Time to implant loss was considered for both groups, and Kaplan-Meier curves were created and compared by the log-rank test.

RESULTS

The IM exhibited the best performance in predicting implant loss risk [AUC = 0.90, 95% confidence interval (CI) 0.84-0.95], followed by the DL model (AUC = 0.87, 95% CI 0.80-0.92) and the CM (AUC = 0.72, 95% CI 0.63-0.79).

CONCLUSION

Our study offers preliminary evidence that both the DL model and IM performed well in predicting implant fate within 5 years and thus may greatly facilitate implant practitioners in assessing preoperative risks.

Decoding the dopamine transporter imaging for the differential diagnosis of parkinsonism using deep learning

Purpose

This work attempts to decode the discriminative information in dopamine transporter (DAT) imaging using deep learning for the differential diagnosis of parkinsonism.

Methods

This study involved 1017 subjects who underwent DAT PET imaging ([¹¹C]CFT) including 43 healthy subjects and 974 parkinsonian patients with idiopathic Parkinson’s disease (IPD), multiple system atrophy (MSA) or progressive supranuclear palsy (PSP). We developed a 3D deep convolutional neural network to learn distinguishable DAT features for the differential diagnosis of parkinsonism. A full-gradient saliency map approach was employed to investigate the functional basis related to the decision mechanism of the network. Furthermore, deep-learning-guided radiomics features and quantitative analysis were compared with their conventional counterparts to further interpret the performance of deep learning.

Results

The proposed network achieved area under the curve of 0.953 (sensitivity 87.7%, specificity 93.2%), 0.948 (sensitivity 93.7%, specificity 97.5%), and 0.900 (sensitivity 81.5%, specificity 93.7%) in the cross-validation, together with sensitivity of 90.7%, 84.1%, 78.6% and specificity of 88.4%, 97.5% 93.3% in the blind test for the differential diagnosis of IPD, MSA and PSP, respectively. The saliency map demonstrated the most contributed areas determining the diagnosis located at parkinsonism-related regions, e.g., putamen, caudate and midbrain. The deep-learning-guided binding ratios showed significant differences among IPD, MSA and PSP groups (P < 0.001), while the conventional putamen and caudate binding ratios had no significant difference between IPD and MSA (P = 0.24 and P = 0.30). Furthermore, compared to conventional radiomics features, there existed average above 78.1% more deep-learning-guided radiomics features that had significant differences among IPD, MSA and PSP.

Conclusion

This study suggested the developed deep neural network can decode in-depth information from DAT and showed potential to assist the differential diagnosis of parkinsonism. The functional regions supporting the diagnosis decision were generally consistent with known parkinsonian pathology but provided more specific guidance for feature selection and quantitative analysis.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00259-022-05804-x.