Deep learning with transfer learning in pathology. Case study: classification of basal cell carcinoma

Area of Science:

• Dermatopathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Basal cell carcinoma (BCC) subtype classification presents diagnostic challenges for pathologists.
• Deep learning (DL) and transfer learning offer powerful image classification capabilities.

Purpose of the Study:

• To design and evaluate a DL-based software for classifying 10 BCC subtypes.
• To assess the performance of transfer learning using established image classification networks.

Main Methods:

• A DL convolution-based software was developed using transfer learning from AlexNet, GoogLeNet, and ResNet-18.
• Three pathologists labeled 2249 BCC patches; 90% for training, 10% for testing.
• Networks were trained on 100 independent sequences and validated on a separate 50-image dataset.

Main Results:

• The software achieved a mean accuracy of 82.53%, sensitivity of 72.52%, specificity of 97.94%, and AUC of 0.99.
• Classification accuracies across different networks were similar, indicating optimal performance on the dataset.
• Software validation demonstrated results comparable to expert pathologist agreement.

Conclusions:

• The developed DL software shows significant promise for classifying BCC subtypes from histological images.
• This AI tool can assist pathologists in diagnosis and medical education.
• Further development could enhance the utility of this software in clinical settings.