Comparative analysis of deep learning architectures and ensemble methods for semantic segmentation of breast cancer histopathological images

Abstract

This work presents a comparative study of deep learning approaches for semantic segmentation of breast cancer histopathological images using the BCSS dataset. The analysis focuses on three architectures — U-Net with ResNet34, DeepLabV3 with EfficientNet-B3, and U-Net with ResNet50 — together with preprocessing, data augmentation, and class-imbalance handling through a combined weighted Cross-Entropy and Dice loss. To improve robustness, Test Time Augmentation and a Soft Voting ensemble were also evaluated. The individual models achieved strong segmentation quality, with mean Dice scores of 0.7935 for U-Net ResNet34, 0.7852 for both DeepLabV3 EfficientNet-B3 and U-Net ResNet50. The ensemble further improved performance, reaching a mean Dice score of 0.8128 and the best overall balance across tissue classes. The most accurately segmented classes were Tumor and Necrosis, while Lymphocytic regions remained the most challenging due to their variability and class imbalance. The results show that combining complementary architectures with TTA-based ensembling can improve segmentation stability and accuracy in histopathological image analysis.

Speaker

Igor N. Tyrin
Samara National Research University
Russia

Discussion

Ask question