To develop and validate a weakly supervised deep learning model for classifying infiltrative basal cell carcinoma (BCC) and poorly differentiated cutaneous squamous cell carcinoma (cSCC) using whole-slide histopathology images, aiming to enhance diagnostic accuracy in clinical settings.
Key Findings:
The model achieved an AUC of 1.0 with 100% accuracy, sensitivity, and specificity on the internal test set, indicating exceptional performance in distinguishing between BCC and cSCC.
In the Queensland cohort, the model maintained an AUC of 1.0 with 90% accuracy at a fixed threshold, demonstrating robustness in external validation.
In the COBRA cohort, the model's AUC was 0.92, with improved accuracy after threshold adjustment using Youden’s J statistic, highlighting the need for calibration in diverse datasets.
Interpretation:
The findings indicate that weakly supervised deep learning can accurately classify challenging subtypes of BCC and cSCC, suggesting significant potential for clinical deployment, although careful calibration and domain adaptation are necessary for reliable use across different institutions.
Limitations:
The in-house data set was monocentric, potentially limiting generalizability to broader populations.
External cohorts were small and lacked detailed subtype annotations, which may affect the robustness of the findings.
Variations in diagnostic subtypes, image resolution, and file formats may impact model performance, necessitating further validation.
Conclusion:
The study demonstrates the potential of deep learning models in dermatopathology, emphasizing the need for calibration and adaptation for broader clinical use, which could significantly enhance diagnostic accuracy.
Chemsex at the pharmacy counter. Gut bacteria tracking helmet impacts. PMD predicting psychiatric illness bidirectionally. This week's research keeps landing in the same uncomfortable place: medicine is improvising.
