A deep learning model for histopathological diagnosis of actinic keratosis: a diagnostic case control accuracy study.

BACKGROUND

Actinic keratosis (AK) is a precancerous skin lesion with the potential to progress into squamous cell carcinoma (SCC), with an overall prevalence of 14%. Although AK is not routinely biopsied, it represents a significant portion of dermatopathology cases. Advances in histopathology now allow for the digitization of slides using whole slide imaging (WSI). While deep learning models (DLMs) have shown promise in various medical fields, their application in dermatopathology remains limited.

METHODS

Building on prior research, we developed a U-Net-based DLM to detect AK in histopathological slides. A total of 371 cases were used to develop the DLM, with two optimal classification thresholds identified for different tasks.

RESULTS

In the test cohort, the DLM achieved an overall accuracy of 98.9% (98.7-99.1) at the patch level and an intersection over union (IoU) of 78.8%. At the WSI level, the DLM reached an accuracy of 96.8% (93.3-98.5) and an IoU of 67.7%. Additionally, 626 remaining cases were evaluated in a Diagnostic Case-Control-Accuracy Study, where the DLM's performance on AK slides and negative controls was compared to the gold standard, achieving an accuracy of 97.6% (96.4-98.5).

CONCLUSIONS

Our findings demonstrate the potential of DLMs to reliably detect AK in routine dermatopathology, suggesting their future impact as technology continues to advance.