Classifying breast intraductal proliferative lesions via a knowledge distillation framework using convolutional neural network-based nuclei-segmentation-assisted classification (KDCNN-NSAC)
Background and objective: Diagnosis of breast intraductal proliferative lesions (BIDPLs) in hematoxylin-eosin (HE) images remains a time-consuming and intractable topic because of subjective processes and subtle morphological differences. Convolutional neural networks (CNNs) show great potential for...
Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-05-01
|
| Series: | Heliyon |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2405844025011375 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Background and objective: Diagnosis of breast intraductal proliferative lesions (BIDPLs) in hematoxylin-eosin (HE) images remains a time-consuming and intractable topic because of subjective processes and subtle morphological differences. Convolutional neural networks (CNNs) show great potential for providing objective analysis strategies for HE images. In this study, we proposed a novel knowledge distillation (KD) framework using CNN-based nuclei segmentation-assisted classification (KDCNN-NSAC). Methods: The diagnosis of BIDPLs is treated as multiple class classification tasks in the BReAst Carcinoma Subtyping dataset. The KDCNN-NSAC fully leveraged the epithelial and stromal nuclei-level features in training phases and performed region-of-interest (ROI)-level classifications in predicting phases. Then, the whole slide image (WSI) was diagnosed based on the risk ratings of the ROIs within it, instead of processing a WSI. Results: The principal results showed that in ROI-level classifications, KDCNN-NSAC outperformed the state-of-the-art methods for 7-class classification with an average F1 score of 63.26 % and achieves F1 score of 98.36 % and 94.21 %, respectively, in distinguishing BIDPLs from invasive cancer and normal tissue. The WSI-level predictions obtained a high degree of consistency with the pathologists’ annotation (kappa value of 0.88). Ablation experiments showed that nuclei segmentation and classification components improve the performance of the baseline model in KDCNN-NSAC by 3 %. Conclusions: The KDCNN-NSAC makes the model focus on important cellular information and predicts the WSI in accordance with the pathologists’ diagnostic thinking, thus improving model explainability. Moreover, the introduce of KDCNN-NSAC will help achieve superior performance in diagnosing BIDPLs. |
|---|---|
| ISSN: | 2405-8440 |