Rapid and precise intraoperative metastasis confirmation using frozen section sentinel lymph node biopsies is crucial for surgical decisions. However, time constraints hinder accurate pathologist diagnosis. Deep learning offers a potential solution, but limited high-quality labeled datasets pose a challenge. This study investigated transfer learning from the CAMELYON16 dataset to improve a convolutional neural network (CNN)-based model's performance on a frozen section dataset (N=297). Results showed that CAMELYON16-based models, even with limited training data, achieved significantly higher AUCs compared to scratch- and ImageNet-based models. This highlights the effectiveness of transfer learning for enhancing tumor classification in frozen section datasets.

Young-Gon Kim, Sungchul Kim, Cristina Eunbee Cho, In Hye Song, Hee Jin Lee, Soomin Ahn, So Yeon Park, Gyungyub Gong, Namkug Kim