A Lightweight Deep Learning Model for Breast Cancer Segmentation on Small Datasets
Xikai Luo, Tianyu Ma, Zhi Fan, Danqing Yin, Jiahui Yu, Yingke Xu · 2024
Breast cancer, a disease with a relatively high incidence, is typically diagnosed by histopathological analysis of microscopic images. However, this manual process suffers from low efficiency and high expertise requirements. Deep learning-based image segmentation offers the potential to automate the identification and delineation of breast cancer lesion areas in histopathological images, thereby improving diagnostic efficiency and lowering the expertise threshold. Existing deep learning models for this task typically require large annotated datasets, posing practical limitations. This paper proposes a convolutional neural network (CNN) based model that can learn effective feature representations from smaller datasets. The model adopts an encoder-decoder architecture, drawing on the deep-shallow feature fusion strategy of DeepLabv3+. By incorporating a Convolutional Block Attention Module (CBAM) and a double-layer convolution-based feature fusion module, the proposed model improves feature extraction capabilities with limited data and mitigates overfitting. Evaluation of the proposed model on the CAMELYON16 dataset demonstrates its ability to extract discriminative features and accurately segment breast cancer lesion areas, even with a relatively small training set. These results highlight the potential of the approach to address the data scarcity challenge in histopathology-based breast cancer diagnosis, paving the way for more accessible and efficient computer-aided diagnosis systems.