Poster Presentation Australasian Society for Dermatology Research Annual Scientific Meeting 2024

Development and Validation of a Deep Learning Model for Predicting Sentinel Lymph Node Metastasis in Melanoma from Histopathological Slides (#70)

Takashi Okamoto 1 , Masataka Kawai 2 , Minh-Khang Le 2 , Shinji Shimada 1 , Tatsuyoshi Kawamura 1
  1. Department of Dermatology, University of Yamanashi, Yamanashi, Japan
  2. Department of Pathology, University of Yamanashi, Yamanashi, Japan

Melanoma originates from the melanocyte, which is the skin pigment-producing cell. Previous histopathology-based deep learning (DL) studies of melanoma focused on diagnostic problems while fewer studies constructed DL models to estimate the clinicopathological behaviors of these tumors. We aimed to construct a DL model to predict the sentinel lymph node metastasis (SLNM) of melanoma, using histopathological patches. Seventy-four, including 35 SLNM-negative and 39 SLNM-positive, melanoma cases were retrieved in the period of 2004-2022 from the University of Yamanashi Hospital (UYH) and divided into train (n=59) and test (n=15) datasets. We constructed a DL model to predict SLNM. This model was trained and tested in train and test datasets, respectively. Patient-level predictions were calculated by averaging 200 patch-level scores of the corresponding slide. The main evaluation metric was the area under the curve (AUC) of receiver operating characteristics (ROC) analysis. The patch-level/patient-level AUCs of our model in train and test sets were 0.75/0.85 and 0.68/0.73. In predicting distant metastasis, AUCs of this model were 0.89 and 0.92 in train and test sets. Dividing by the median score, high-score patients had significantly worse progression-free survival (p=0.059). Histopathology analysis showed that melanin depositions, dense tumor sheets, and nuclear atypia were important morphologic factors attributed to the “positive” predictions of our model. By using DL models, histopathology of melanoma can be used to predict SLNM, which may help clinical decisions. However, more samples are needed to train a more robust model for clinical uses.