Multiple instance learning for computer aided detection and diagnosis of gastric cancer with dual-energy CT imaging

Artificial intelligence in screening and diagnosis of surgical diseases: A narrative review

Artificial intelligence (AI) is playing an increasing role in several fields of medicine. It is also gaining popularity among surgeons as a valuable screening and diagnostic tool for many conditions such as benign and malignant colorectal, gastric, thyroid, parathyroid, and breast disorders. In the literature, there is no review that groups together the various application domains of AI when it comes to the screening and diagnosis of main surgical diseases. The aim of this review is to describe the use of AI in these settings. We performed a literature review by searching PubMed, Web of Science, Scopus, and Embase for all studies investigating the role of AI in the surgical setting, published between January 01, 2000, and June 30, 2023. Our focus was on randomized controlled trials (RCTs), meta-analysis, systematic reviews, and observational studies, dealing with large cohorts of patients. We then gathered further relevant studies from the reference list of the selected publications. Based on the studies reviewed, it emerges that AI could strongly enhance the screening efficiency, clinical ability, and diagnostic accuracy for several surgical conditions. Some of the future advantages of this technology include implementing, speeding up, and improving the automaticity with which AI recognizes, differentiates, and classifies the various conditions.

Weakly Supervised MRI Slice-Level Deep Learning Classification of Prostate Cancer Approximates Full Voxel- and Slice-Level Annotation: Effect of Increasing Training Set Size

BACKGROUND

Weakly supervised learning promises reduced annotation effort while maintaining performance.

PURPOSE

To compare weakly supervised training with full slice-wise annotated training of a deep convolutional classification network (CNN) for prostate cancer (PC).

STUDY TYPE

Retrospective.

SUBJECTS

One thousand four hundred eighty-nine consecutive institutional prostate MRI examinations from men with suspicion for PC (65 ± 8 years) between January 2015 and November 2020 were split into training (N = 794, enriched with 204 PROSTATEx examinations) and test set (N = 695).

FIELD STRENGTH/SEQUENCE

1.5 and 3T, T2-weighted turbo-spin-echo and diffusion-weighted echo-planar imaging.

ASSESSMENT

Histopathological ground truth was provided by targeted and extended systematic biopsy. Reference training was performed using slice-level annotation (SLA) and compared to iterative training utilizing patient-level annotations (PLAs) with supervised feedback of CNN estimates into the next training iteration at three incremental training set sizes (N = 200, 500, 998). Model performance was assessed by comparing specificity at fixed sensitivity of 0.97 [254/262] emulating PI-RADS ≥ 3, and 0.88-0.90 [231-236/262] emulating PI-RADS ≥ 4 decisions.

STATISTICAL TESTS

Receiver operating characteristic (ROC) and area under the curve (AUC) was compared using DeLong and Obuchowski test. Sensitivity and specificity were compared using McNemar test. Statistical significance threshold was P = 0.05.

RESULTS

Test set (N = 695) ROC-AUC performance of SLA (trained with 200/500/998 exams) was 0.75/0.80/0.83, respectively. PLA achieved lower ROC-AUC of 0.64/0.72/0.78. Both increased performance significantly with increasing training set size. ROC-AUC for SLA at 500 exams was comparable to PLA at 998 exams (P = 0.28). ROC-AUC was significantly different between SLA and PLA at same training set sizes, however the ROC-AUC difference decreased significantly from 200 to 998 training exams. Emulating PI-RADS ≥ 3 decisions, difference between PLA specificity of 0.12 [51/433] and SLA specificity of 0.13 [55/433] became undetectable (P = 1.0) at 998 exams. Emulating PI-RADS ≥ 4 decisions, at 998 exams, SLA specificity of 0.51 [221/433] remained higher than PLA specificity at 0.39 [170/433]. However, PLA specificity at 998 exams became comparable to SLA specificity of 0.37 [159/433] at 200 exams (P = 0.70).

DATA CONCLUSION

Weakly supervised training of a classification CNN using patient-level-only annotation had lower performance compared to training with slice-wise annotations, but improved significantly faster with additional training data.

EVIDENCE LEVEL

3 TECHNICAL EFFICACY: Stage 2.

Diagnostic accuracy of radiomics-based machine learning for neoadjuvant chemotherapy response and survival prediction in gastric cancer patients: A systematic review and meta-analysis

BACKGROUND

In recent years, researchers have explored the use of radiomics to predict neoadjuvant chemotherapy outcomes in gastric cancer (GC). Yet, a lingering debate persists regarding the accuracy of these predictions. Against this backdrop, this study was conducted to examine the accuracy of radiomics in predicting the response to neoadjuvant chemotherapy in GC patients.

METHODS

An exhaustive search of relevant studies was conducted in PubMed, Cochrane, Embase, and Web of Science databases up to February 21, 2023. The radiomics quality scoring (RQS) tool was employed to assess study quality. Tumor response to neoadjuvant chemotherapy and survival outcomes were examined as outcome measures.

RESULTS

Fourteen studies involving 3,373 GC patients who had received neoadjuvant chemotherapy were incorporated in our meta-analysis. The mean RQS score across all studies was 36.3%, ranging between 0 and 63.9%. On the validation cohort, when the modeling variables were restricted to radiomic features alone, the predictive performance was characterized by a c-index of 0.750 (95% CI: 0.710-0.790), with a sensitivity of 0.67 (95% CI: 0.58-0.75) and a specificity of 0.77 (95% CI: 0.69-0.84) for the prediction of neoadjuvant chemotherapy response. When clinical data was integrated with radiomic features as modeling variables, the predictive performance improved, yielding a c-index of 0.814 (95% CI: 0.780-0.847), a sensitivity of 0.78 [95% CI: 0.70-0.84], and a specificity of 0.73 [95% CI: 0.67-0.79].

CONCLUSIONS

Radiomics holds promise to noninvasively predict neoadjuvant chemotherapy response and survival outcomes among patients with locally advanced GC. Additionally, we underscore the need for future multicenter studies and the development of imaging-sourced tools for risk stratification in this patient population.

Artificial intelligence in theranostics of gastric cancer, a review

Gastric cancer (GC) is one of the commonest cancers with high morbidity and mortality in the world. How to realize precise diagnosis and therapy of GC owns great clinical requirement. In recent years, artificial intelligence (AI) has been actively explored to apply to early diagnosis and treatment and prognosis of gastric carcinoma. Herein, we review recent advance of AI in early screening, diagnosis, therapy and prognosis of stomach carcinoma. Especially AI combined with breath screening early GC system improved 97.4 % of early GC diagnosis ratio, AI model on stomach cancer diagnosis system of saliva biomarkers obtained an overall accuracy of 97.18 %, specificity of 97.44 %, and sensitivity of 96.88 %. We also discuss concept, issues, approaches and challenges of AI applied in stomach cancer. This review provides a comprehensive view and roadmap for readers working in this field, with the aim of pushing application of AI in theranostics of stomach cancer to increase the early discovery ratio and curative ratio of GC patients.

Accurate preoperative staging and HER2 status prediction of gastric cancer by the deep learning system based on enhanced computed tomography

Purpose

To construct the deep learning system (DLS) based on enhanced computed tomography (CT) images for preoperative prediction of staging and human epidermal growth factor receptor 2 (HER2) status in gastric cancer patients.

Methods

The raw enhanced CT image dataset consisted of CT images of 389 patients in the retrospective cohort, The Cancer Imaging Archive (TCIA) cohort, and the prospective cohort. DLS was developed by transfer learning for tumor detection, staging, and HER2 status prediction. The pre-trained Yolov5, EfficientNet, EfficientNetV2, Vision Transformer (VIT), and Swin Transformer (SWT) were studied. The tumor detection and staging dataset consisted of 4860 enhanced CT images and annotated tumor bounding boxes. The HER2 state prediction dataset consisted of 38900 enhanced CT images.

Results

The DetectionNet based on Yolov5 realized tumor detection and staging and achieved a mean Average Precision (IoU=0.5) (mAP_0.5) of 0.909 in the external validation cohort. The VIT-based PredictionNet performed optimally in HER2 status prediction with the area under the receiver operating characteristics curve (AUC) of 0.9721 and 0.9995 in the TCIA cohort and prospective cohort, respectively. DLS included DetectionNet and PredictionNet had shown excellent performance in CT image interpretation.

Conclusion

This study developed the enhanced CT-based DLS to preoperatively predict the stage and HER2 status of gastric cancer patients, which will help in choosing the appropriate treatment to improve the survival of gastric cancer patients.

Hybrid Deep Learning Model for Endoscopic Lesion Detection and Classification Using Endoscopy Videos

In medical imaging, the detection and classification of stomach diseases are challenging due to the resemblance of different symptoms, image contrast, and complex background. Computer-aided diagnosis (CAD) plays a vital role in the medical imaging field, allowing accurate results to be obtained in minimal time. This article proposes a new hybrid method to detect and classify stomach diseases using endoscopy videos. The proposed methodology comprises seven significant steps: data acquisition, preprocessing of data, transfer learning of deep models, feature extraction, feature selection, hybridization, and classification. We selected two different CNN models (VGG19 and Alexnet) to extract features. We applied transfer learning techniques before using them as feature extractors. We used a genetic algorithm (GA) in feature selection, due to its adaptive nature. We fused selected features of both models using a serial-based approach. Finally, the best features were provided to multiple machine learning classifiers for detection and classification. The proposed approach was evaluated on a personally collected dataset of five classes, including gastritis, ulcer, esophagitis, bleeding, and healthy. We observed that the proposed technique performed superbly on Cubic SVM with 99.8% accuracy. For the authenticity of the proposed technique, we considered these statistical measures: classification accuracy, recall, precision, False Negative Rate (FNR), Area Under the Curve (AUC), and time. In addition, we provided a fair state-of-the-art comparison of our proposed technique with existing techniques that proves its worthiness.

Dual-Energy Computed Tomography-Based Radiomics to Predict Peritoneal Metastasis in Gastric Cancer

Objective

To develop and validate a dual-energy computed tomography (DECT) derived radiomics model to predict peritoneal metastasis (PM) in patients with gastric cancer (GC).

Methods

This retrospective study recruited 239 GC (non-PM = 174, PM = 65) patients with histopathological confirmation for peritoneal status from January 2015 to December 2019. All patients were randomly divided into a training cohort (n = 160) and a testing cohort (n = 79). Standardized iodine-uptake (IU) images and 120-kV-equivalent mixed images (simulating conventional CT images) from portal-venous and delayed phases were used for analysis. Two regions of interest (ROIs) including the peritoneal area and the primary tumor were independently delineated. Subsequently, 1691 and 1226 radiomics features were extracted from the peritoneal area and the primary tumor from IU and mixed images on each phase. Boruta and Spearman correlation analysis were used for feature selection. Three radiomics models were established, including the R_IU model for IU images, the R_MIX model for mixed images and the combined radiomics model (the R_comb model). Random forest was used to tune the optimal radiomics model. The performance of the clinical model and human experts to assess PM was also recorded.

Results

Fourteen and three radiomics features with low redundancy and high importance were extracted from the IU and mixed images, respectively. The R_IU model showed significantly better performance to predict PM than the R_MIX model in the training cohort (AUC, 0.981 vs. 0.917, p = 0.034). No improvement was observed in the R_comb model (AUC = 0.967). The R_IU model was the optimal radiomics model which showed no overfitting in the testing cohort (AUC = 0.967, p = 0.528). The R_IU model demonstrated significantly higher predictive value on peritoneal status than the clinical model and human experts in the testing cohort (AUC, 0.785, p = 0.005; AUC, 0.732, p <0.001, respectively).

Conclusion

DECT derived radiomics could serve as a non-invasive and easy-to-use biomarker to preoperatively predict PM for GC, providing opportunity for those patients to tailor appropriate treatment.

Artificial intelligence in gastric cancer: Application and future perspectives

Gastric cancer is the fourth leading cause of cancer-related mortality across the globe, with a 5-year survival rate of less than 40%. In recent years, several applications of artificial intelligence (AI) have emerged in the gastric cancer field based on its efficient computational power and learning capacities, such as image-based diagnosis and prognosis prediction. AI-assisted diagnosis includes pathology, endoscopy, and computerized tomography, while researchers in the prognosis circle focus on recurrence, metastasis, and survival prediction. In this review, a comprehensive literature search was performed on articles published up to April 2020 from the databases of PubMed, Embase, Web of Science, and the Cochrane Library. Thereby the current status of AI-applications was systematically summarized in gastric cancer. Moreover, future directions that target this field were also analyzed to overcome the risk of overfitting AI models and enhance their accuracy as well as the applicability in clinical practice.

Extracting 2D weak labels from volume labels using multiple instance learning in CT hemorrhage detection

Multiple instance learning (MIL) is a supervised learning methodology that aims to allow models to learn instance class labels from bag class labels, where a bag is defined to contain multiple instances. MIL is gaining traction for learning from weak labels but has not been widely applied to 3D medical imaging. MIL is well-suited to clinical CT acquisitions since (1) the highly anisotropic voxels hinder application of traditional 3D networks and (2) patch-based networks have limited ability to learn whole volume labels. In this work, we apply MIL with a deep convolutional neural network to identify whether clinical CT head image volumes possess one or more large hemorrhages (> 20cm3), resulting in a learned 2D model without the need for 2D slice annotations. Individual image volumes are considered separate bags, and the slices in each volume are instances. Such a framework sets the stage for incorporating information obtained in clinical reports to help train a 2D segmentation approach. Within this context, we evaluate the data requirements to enable generalization of MIL by varying the amount of training data. Our results show that a training size of at least 400 patient image volumes was needed to achieve accurate per-slice hemorrhage detection. Over a five-fold cross-validation, the leading model, which made use of the maximum number of training volumes, had an average true positive rate of 98.10%, an average true negative rate of 99.36%, and an average precision of 0.9698. The models have been made available along with source code1 to enabled continued exploration and adaption of MIL in CT neuroimaging.

Functional Site Discovery From Incomplete Training Data: A Case Study With Nucleic Acid-Binding Proteins

Function annotation efforts provide a foundation to our understanding of cellular processes and the functioning of the living cell. This motivates high-throughput computational methods to characterize new protein members of a particular function. Research work has focused on discriminative machine-learning methods, which promise to make efficient, de novo predictions of protein function. Furthermore, available function annotation exists predominantly for individual proteins rather than residues of which only a subset is necessary for the conveyance of a particular function. This limits discriminative approaches to predicting functions for which there is sufficient residue-level annotation, e.g., identification of DNA-binding proteins or where an excellent global representation can be divined. Complete understanding of the various functions of proteins requires discovery and functional annotation at the residue level. Herein, we cast this problem into the setting of multiple-instance learning, which only requires knowledge of the protein’s function yet identifies functionally relevant residues and need not rely on homology. We developed a new multiple-instance leaning algorithm derived from AdaBoost and benchmarked this algorithm against two well-studied protein function prediction tasks: annotating proteins that bind DNA and RNA. This algorithm outperforms certain previous approaches in annotating protein function while identifying functionally relevant residues involved in binding both DNA and RNA, and on one protein-DNA benchmark, it achieves near perfect classification.

A survey on Barrett's esophagus analysis using machine learning

This work presents a systematic review concerning recent studies and technologies of machine learning for Barrett's esophagus (BE) diagnosis and treatment. The use of artificial intelligence is a brand new and promising way to evaluate such disease. We compile some works published at some well-established databases, such as Science Direct, IEEEXplore, PubMed, Plos One, Multidisciplinary Digital Publishing Institute (MDPI), Association for Computing Machinery (ACM), Springer, and Hindawi Publishing Corporation. Each selected work has been analyzed to present its objective, methodology, and results. The BE progression to dysplasia or adenocarcinoma shows a complex pattern to be detected during endoscopic surveillance. Therefore, it is valuable to assist its diagnosis and automatic identification using computer analysis. The evaluation of the BE dysplasia can be performed through manual or automated segmentation through machine learning techniques. Finally, in this survey, we reviewed recent studies focused on the automatic detection of the neoplastic region for classification purposes using machine learning methods.

A ℓ2, 1 norm regularized multi-kernel learning for false positive reduction in Lung nodule CAD

OBJECTIVE

The aim of this paper is to describe a novel algorithm for False Positive Reduction in lung nodule Computer Aided Detection(CAD).

METHODS

In this paper, we describes a new CT lung CAD method which aims to detect solid nodules. Specially, we proposed a multi-kernel classifier with a ℓ_{2, 1} norm regularizer for heterogeneous feature fusion and selection from the feature subset level, and designed two efficient strategies to optimize the parameters of kernel weights in non-smooth ℓ_{2, 1} regularized multiple kernel learning algorithm. The first optimization algorithm adapts a proximal gradient method for solving the ℓ_{2, 1} norm of kernel weights, and use an accelerated method based on FISTA; the second one employs an iterative scheme based on an approximate gradient descent method.

RESULTS

The results demonstrates that the FISTA-style accelerated proximal descent method is efficient for the ℓ_{2, 1} norm formulation of multiple kernel learning with the theoretical guarantee of the convergence rate. Moreover, the experimental results demonstrate the effectiveness of the proposed methods in terms of Geometric mean (G-mean) and Area under the ROC curve (AUC), and significantly outperforms the competing methods.

CONCLUSIONS

The proposed approach exhibits some remarkable advantages both in heterogeneous feature subsets fusion and classification phases. Compared with the fusion strategies of feature-level and decision level, the proposed ℓ_{2, 1} norm multi-kernel learning algorithm is able to accurately fuse the complementary and heterogeneous feature sets, and automatically prune the irrelevant and redundant feature subsets to form a more discriminative feature set, leading a promising classification performance. Moreover, the proposed algorithm consistently outperforms the comparable classification approaches in the literature.

Multiple-Instance Learning for Anomaly Detection in Digital Mammography

This paper describes a computer-aided detection and diagnosis system for breast cancer, the most common form of cancer among women, using mammography. The system relies on the Multiple-Instance Learning (MIL) paradigm, which has proven useful for medical decision support in previous works from our team. In the proposed framework, breasts are first partitioned adaptively into regions. Then, features derived from the detection of lesions (masses and microcalcifications) as well as textural features, are extracted from each region and combined in order to classify mammography examinations as "normal" or "abnormal". Whenever an abnormal examination record is detected, the regions that induced that automated diagnosis can be highlighted. Two strategies are evaluated to define this anomaly detector. In a first scenario, manual segmentations of lesions are used to train an SVM that assigns an anomaly index to each region; local anomaly indices are then combined into a global anomaly index. In a second scenario, the local and global anomaly detectors are trained simultaneously, without manual segmentations, using various MIL algorithms (DD, APR, mi-SVM, MI-SVM and MILBoost). Experiments on the DDSM dataset show that the second approach, which is only weakly-supervised, surprisingly outperforms the first approach, even though it is strongly-supervised. This suggests that anomaly detectors can be advantageously trained on large medical image archives, without the need for manual segmentation.