200 Deep learning model to predict sentinel lymph node status in melanoma patients

The application of artificial intelligence (AI) to medicine is being studied in various fields. However, most them have been developed for "diagnosis" based on learning from clinical images. Our ultimate goal develop AI "treatment responsiveness" and "patient prognosis prediction" as a new strategy. treatment malignant melanoma has greatly advanced by immune checkpoint inhibitors. there are no deep learning-based histopathological biomarkers melanoma. Recently, we trying an that can predict (ICI response, overall survival rate, etc.) using images (WSI: Whole Slide Imaging) We first focused sentinel lymph nodes (SLNs). SLNs status important prognostic factor patients. Therefore, the aim this study digital biomarker noninvasively node metastasis WSI tissue. use around 400 skin cutaneous (SKCM) samples (WSI LN status) Cancer Genome Atlas (TCGA) database training validation dataset. Model was performed imagenet pretrained convnet followed attention-pooling layer. area under ROC curve (AUROC) used evaluate accuracy. hyperparameters with largest AUROC were searched 5-fold cross training. best prediction 0.65. This model currently undergoing dataset University Yamanashi (in-house dataset) improve its results indicate histological features primary some extent metastasis.