Ultrasonographic Phenotypes and Segmentation Model Comparison for Primary Breast Lesions and Axillary Lymph Nodes in Triple-negative Breast Cancer.
Authors
Affiliations (2)
Affiliations (2)
- Department of Ultrasound, The First Affiliated Hospital of Dalian Medical University, Dalian, Liaoning 116011, China (J.L., X.B., S.X. L.S.).
- Department of Ultrasound, The First Affiliated Hospital of Dalian Medical University, Dalian, Liaoning 116011, China (J.L., X.B., S.X. L.S.). Electronic address: [email protected].
Abstract
To describe clinicopathologic and ultrasonographic heterogeneity across immunohistochemistry-based FUSCC surrogate subtypes in triple-negative breast cancer (TNBC) and compare 3 automatic segmentation models for primary breast lesions and axillary lymph nodes (ALNs) on two-dimensional ultrasound. This single-center retrospective study included 102 TNBC cases with breast lesion ultrasound images and manual masks; 54 also had evaluable ALN images and masks. Subtype analyses included classified luminal androgen receptor (LAR), immunomodulatory (IM), and basal-like immune-suppressed (BLIS) cases. Fixed splits were 72/15/15 for breast lesions and 38/8/8 for ALNs. UNet++, DeepLabV3+, and nnU-Net were compared using Dice, IoU, Precision, Recall, and Specificity. Bootstrap confidence intervals and paired Wilcoxon signed-rank tests were used. Bland-Altman analysis assessed ultrasound-pathology size agreement in 92 cases. Among 84 classified cases, 35 were LAR, 16 IM, and 33 BLIS. LAR cases were older, and posterior acoustic patterns differed across subtypes. Breast lesion Dice scores were 0.7866, 0.8698, and 0.8704 for UNet++, DeepLabV3+, and nnU-Net; ALN Dice scores were 0.6005, 0.8007, and 0.8097. DeepLabV3+ significantly outperformed UNet++ in both tasks, whereas DeepLabV3+ and nnU-Net did not differ significantly. Mean ultrasound-pathology size bias was -1.28 mm. Classified TNBC cases showed clinicopathologic and ultrasonographic heterogeneity. DeepLabV3+ and nnU-Net showed numerically favorable segmentation performance in this exploratory fixed-split cohort and may be considered candidate region-of-interest (ROI)-generation approaches, pending validation in larger independent datasets.