Abstract

This study aimed to evaluate the effectiveness of computer vision methods in the diagnosis of abdominal pathologies. The methodology included a search of databases covering publications from 2019 to September 2025. It described the application of convolutional neural network architectures, segmentation methods, and multimodal approaches to medical image analysis. The reviewed data showed that DenseNet-201 and EfficientNetV2M achieved accuracies of 97.1-99.4% in classifying endoscopic lesions, outperforming AlexNet and ResNet-101. Convolutional neural network-based systems demonstrated diagnostic performance comparable to that of experts (diagnostic odds ratio 1.03), with sensitivity of 84.8-94.3% and an area under the curve of 0.95-0.98 in the detection of colorectal polyps. Multimodal approaches, including a hierarchical fully convolutional neural network for positron emission tomography-computed tomography and the ENDOANGEL-MM system for endoscopy, achieved a 21% increase in accuracy relative to unimodal models. Automated segmentation systems reduced analysis time by 92.8% while maintaining Dice similarity coefficients of 0.936-0.965 and reducing inter-expert variability from 7.78% to 4.10%. In critical diagnostics, models for detecting parenchymal organ injuries showed variability in performance across organs (receiver operating characteristic area under the curve – 0.841-0.945; sensitivity – 0.736-0.862). The combined model achieved the highest indicators – accuracy of 0.954 and specificity of 0.966, confirming its suitability for primary screening. Data augmentation and multitask learning methods increased accuracy by 15-17.6% for rare pathologies, while domain randomisation ensured an increase in the Dice similarity coefficient of up to 15.8% on independent data. The practical significance of the results lies in their potential application by healthcare institutions to implement automated diagnostic systems, optimise radiological workflows, and standardise clinical conclusions, thereby reducing time to treatment initiation and decreasing diagnostic errors