Jazarix

Federated learning-based colorectal cancer classification by convolutional neural networks and general visual representation learning

Colorectal cancer is the fourth leading cause of death worldwide, and deep learning (DL) methods can help pathologists classify precancerous and cancerous colorectal lesions more efficiently. However, patient data privacy remains a challenge when training DL models with large medical datasets. This study demonstrates that Federated Learning, which allows model training across multiple sites without compromising privacy, enhances classification performance across four data settings from the MHIST and Chaoyang datasets. The Big Transfer and VGG models achieved the highest AUC scores, showing that federated learning can effectively support collaborative efforts in computational pathology.