Artificial neural networks and pathologists recognize basal cell carcinomas based on different histological patterns.

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Basal cell carcinoma (BCC) is the most common skin tumor, requiring daily analysis of histological slides by pathologists.
• Deep learning and artificial intelligence (AI) offer potential for automated analysis of medical images, including histological slides.
• Automated prescreening of histological whole-slide images (WSIs) can aid in identifying cancerous regions and tumor classification.

Purpose of the Study:

• To implement an accurate and interpretable artificial neural network (ANN) for detecting BCCs in histological WSIs.
• To compare the diagnostic histological features and recognition patterns used by the ANN versus expert pathologists.
• To evaluate the potential of machine learning algorithms to enhance diagnostic precision in digital pathology.

Main Methods:

• An attention-based ANN was trained using 820 WSIs of BCCs to identify tumor regions.
• The diagnostic regions identified by the ANN were compared with pathologists' regions of interest using eye-tracking techniques.
• Performance was evaluated using metrics such as area under the ROC curve, sensitivity, and specificity.

Main Results:

• The ANN demonstrated high accuracy in identifying BCC tumor regions, with an area under the ROC curve of 0.993.
• The network achieved a sensitivity of 0.965 and a specificity of 0.910 in BCC detection.
• Machine learning algorithms utilized significantly different recognition patterns for tumor identification compared to expert pathologists (p < 10^-4).

Conclusions:

• State-of-the-art machine learning techniques can efficiently and interpretably analyze histopathological images, as demonstrated with BCC WSIs.
• Artificial neural networks and machine learning algorithms show promise in enhancing diagnostic accuracy in digital pathology.
• These AI tools may uncover novel classification patterns previously unrecognized by human experts.