508 Enhancing AI-Assisted Melanoma Diagnosis Through a Human–Machine Interactive Framework

Introduction/Objective The lack of high-quality, annotated datasets has been a major barrier to the application of artificial intelligence (AI) in pathology slide diagnosis. Annotating pathology slides is time-consuming, requires the expertise of trained pathologists, and is often prohibitively expensive when scaling to large datasets. Additionally, inter-observer variability can introduce noise and inconsistencies into the training data. To address these challenges, we developed a human–machine interactive program based on the U-Net architecture, enabling rapid and efficient annotation of large volumes of pathology slides. This approach facilitates the generation of high-quality labeled datasets that can be used to train machine learning and deep learning models to assist in accurate and scalable pathology diagnosis. Methods/Case Report Whole slide images of melanoma pathology slides were discretized into tiles and subsequently divided into smaller patches. These patches were clustered to facilitate human selection and annotation. The selected and annotated patches were then used to train a U-Net model, which generated preliminary annotations. These initial annotations were undergone further refinement and used as input for iterative retraining of the U-Net model to improve its performance. Results Using a dataset of whole slide images, we successfully applied our program to rapidly and accurately annotate over 100 melanoma slides. The resulting high-quality annotations were then used to train machine learning models, which demonstrated significantly improved performance on an external validation dataset, highlighting the robustness and generalizability of our approach. Conclusion We have developed a human–machine interactive program capable of rapidly annotating melanoma pathology slides with high accuracy. The annotations generated by this program can be used to train high-quality machine learning models that assist pathologists in improving the diagnostic accuracy of melanoma slides.