Multi-output UNet model for simultaneous segmentation and classification of breast volume sweep imaging ultrasound
Monday, April 8, 2024
12:09pm – 12:19pm
Location: 406
Authors: Donya Khaledyan, University of Rochester Thomas Marini, University of Rochester Kevin Parker, University of Rochester
In addressing the challenge of breast lesion segmentation and classification, especially in resource-constrained rural areas, our proposed CNN architecture integrates UNet-based segmentation with dense classification layers for simultaneous learning. This pioneering model showcases superior performance, excelling in accuracy, speed, and computational efficiency, effectively catering to evolving clinical data and disease classifications. Notably, it introduces a novel joint loss function, allowing shared parameter optimization during training. Leveraging a dataset of 3818 VSI scans, our model demonstrated exceptional classification results (2% loss, 99% accuracy, sensitivity, and specificity) and competitive segmentation metrics (10% Dice loss, 87% F1 score, 85% recall, 99% specificity, 87% Dice coefficient), surpassing existing pre-trained models. The joint segmentation classification model emphasizes efficient feature sharing, compactness, and applicability, significantly enhancing healthcare accessibility in underserved regions, where radiologists are scarce.