AI13 Development and performance evaluation of a deep learning model for the histopathological diagnosis of actinic keratosis: a diagnostic case–control accuracy study
Julius Balkenhol, Maximilian Schmidt, T. Schnauder, Johannes Langhorst, Jean Le’Clerc Arrastia, Daniel Otero Baguer, Georgia Gilbert, Lutz Schmitz, Thomas Dirschka · British Journal of Dermatology · 2025
Abstract Actinic keratoses (AKs) are precancerous skin lesions that have the potential to progress to squamous cell carcinoma. Although AKs are not routinely biopsied, they represent a significant portion of dermatopathology cases. Advances in histopathology now allow for the digitization of slides using whole-slide imaging (WSI). A deep learning model (DLM) is a type of artificial neural network that is capable of solving diagnostic problems by learning hierarchical data representations. DLMs have demonstrated promise in diagnosing breast cancer; however, their use in dermatopathology remains limited (Olsen TG, Jackson BH, Feeser TA et al. Diagnostic performance of deep learning algorithms applied to three common diagnoses in dermatopathology. J Pathol Inform 2018; 9: 32). We aimed to evaluate whether a DLM could assist in the dermatopathological diagnosis of AK. We developed a U-Net-based DLM to detect AK in histopathological samples. Firstly, the DLM was trained on a ‘training cohort’ of 164 cases of AK and 18 controls for 60 epochs using the Adam optimizer, with the goal of aligning its output with expert dermatopathologist annotations. Optimal thresholds for classification were identified using the training model: 0.35 at the patch level to maximize sensitivity while maintaining high specificity, and a more conservative threshold of 0.45 at the WSI level. Next, the DLM was evaluated on a ‘test cohort’ of 105 cases of AK and 84 controls to assess its technical performance at the pixel, patch and WSI levels. Finally, diagnostic performance was evaluated on 815 cases (731 AKs and 84 controls) in a case–control accuracy study. In the test cohort, the DLM achieved 98.9% accuracy and 78.8% intersection over union (IoU) at the patch level, and 96.8% accuracy and 67.7% IoU at the WSI level. The overall diagnostic performance metrics are displayed in the Table. Our findings show early potential of DLMs to reliably detect AK in routine dermatopathology. However, there are several limitations, such as the DLM’s inability to distinguish between different grades of AK or diagnose other pathologies. There are also high costs of digitizing images and integrating the technology into current workflows. Future research is required to assess whether DLMs could be used to reduce dermatopathologists’ workload and provide valuable research opportunities through large-scale data analysis.TableOverall diagnostic performance evaluation of the deep learning model at whole-slide imaging level (total n = 815)True positives717False negatives18True negatives79False positives1Positive predictive valve99.9%Negative predictive valve81.4%Sensitivity97.5%Specificity98.8%Accuracy97.6%F1 score98.7%