Segmentation and classification of radiological images with deep learning approaches: A clinical application

Aim: This study was carried out in three stages. First, the segmentation and classification performances of various deep learning architectures were compared. In the second stage, the most effective algorithms were identified. Finally, two separate clinical decision-support systems were developed using the selected algorithms—one for automatic segmentation of the carotid artery structure from radiological images, and the other for automatic segmentation and classification of breast lesions. Materials and Methods: Two datasets were used: breast ultrasound images and carotid artery ultrasound images. On these datasets, nine deep learning architectures U-Net, SegNet, DeepLab v3+, Mask R-CNN, Attention U-Net, U-Net++, TransUNet, Swin-Unet, and Vision Mamba UNet were compared in terms of segmentation performance. Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics were employed for performance evaluation. Additionally, the breast ultrasound images were classified using nine different CNN models as well as Swin Transformer and Vision Transformer. The segmentation and classification algorithms that achieved the highest performance were then integrated into the web-based clinical applications developed for this thesis. Results: Vision Mamba UNet demonstrated the highest segmentation performance across both datasets, achieving an average DSC of 0.943 and an IoU of 0.896; SwinUNet showed similarly strong performance. For classification, Vision Transformer stood out with an F1 score of 88.7. The web applications based on Vision Mamba UNet and Vision Transformer—for carotid artery segmentation and breast lesion segmentation/classification, respectively—yielded positive feedback in terms of ease of use, speed, and accuracy. Conclusion: The developed web-based applications have the potential to be incorporated into clinical decision workflows, reducing radiologists' workload, accelerating the diagnostic process, and improving diagnostic accuracy. Keywords: Breast lesion segmentation and classification, Carotid artery segmentation, Clinical decision support system, Deep learning.