Enhancing basal cell carcinoma classification in preoperative biopsies via transfer learning with weakly supervised
Johan Björkman1,2, Sigrid Lagerroth2, Jan Siarov2,3
1Department of Physics, Chalmers University of Technology, Gothenburg, Sweden.
BMC Medical Imaging
|May 16, 2025
Summary
Transfer learning significantly improves automated basal cell carcinoma (BCC) classification in biopsies, achieving high accuracy even with limited data. This approach enhances diagnostic precision for this common skin cancer.
Area of Science:
- Computational pathology
- Machine learning in histopathology
- Digital diagnostics
Background:
- Basal cell carcinoma (BCC) is the most prevalent skin cancer globally, posing a substantial healthcare burden.
- High-precision automated BCC diagnostics necessitate large annotated datasets, which are challenging and expensive to acquire.
- This study addresses the need for efficient diagnostic tools by fine-tuning machine learning models for BCC classification.
Purpose of the Study:
- To fine-tune a weakly supervised machine learning model for classifying basal cell carcinoma (BCC) in preoperative punch biopsies.
- To leverage transfer learning to overcome challenges of data scarcity and variability in computational pathology.
- To enhance the generalizability and diagnostic accuracy of automated BCC detection systems.
Main Methods:
- Utilized the Basal Cell Classification (BCCC) dataset of 514 whole slide images (WSIs) of punch biopsies for training, validation, and testing.
- Extracted features using a pretrained simCLR model and processed them with a Vision Transformer, incorporating spatial information via graph formation.
- Evaluated fine-tuned, non-fine-tuned pretrained, and scratch models on BCCC and externally validated on the COBRA dataset (3,588 WSIs).
Main Results:
- The fine-tuned model achieved superior performance, with accuracies of 91.7% (2-class), 82.1% (3-class), and 75.3% (5-class) on the BCCC dataset.
- External validation on the COBRA dataset showed accuracies of 84.9% (2-class) and 70.5% (3-class).
- Ablation studies confirmed the fine-tuned model's superiority, demonstrating significant improvements in mean accuracy over models trained from scratch.
Conclusions:
- Transfer learning effectively enhances model performance, particularly in data-limited scenarios common in computational pathology.
- The study validates the utility of pre-trained models for robust feature extraction in complex histopathological analyses.
- This approach offers a scalable and accurate solution for automated BCC classification, addressing limitations in annotated dataset availability.
