Cross-Organ Investigation of Tumor Histological Features Similarities Using Transfer Learning: A Case Study on Breast and Colorectal Tumors

Abstract

Breast and colorectal cancers are two of the most commonly occurring cancers in the world. Transfer Learning with models pre-trained on ImageNet has been extensively used in the literature in the detection of these two deadly diseases from histopathology images. A limited number of works have investigated cross organ histological similarities using deep learning, which showed the correlation between some breast and colorectal cancer subtypes. In this paper, we focus on a further investigation of the similarities and correlation between breast and colorectal cancers by focusing on the binary benign/malignant classification problem. We conduct a number of experiments with different training and fine-tuning strategies, leveraging transfer learning from pre-trained models. Using the different strategies, a model is trained on a breast cancer histopathology dataset, and tested on two colorectal cancer histopathology datasets. Accordingly, the results demonstrate similarities in benign and malignant tumors across the two organs, with an accuracy reaching as high as 97.33% and 98.80% on the benign and malignant colorectal samples respectively.