Development of a Transfer Learning-Based Interactive Tool for Supporting Tissue-Type Classification in Pathology Images

Abstract

Pathology images are crucial for accurate diagnoses, but accurate tissue classification requires significant expertise and effort. This study developed an interactive classification tool that sequentially learns from user input and incrementally improves through user modifications. The tool employs a random forest to classify Masson trichrome-stained images into ”cell nuclei,” ”collagen,” and ”background,” and can extend its classification capability to new images through transfer learning. The evaluation results confirmed that the tool consistently provides higher accuracy and significantly improves operational efficiency compared to manual classification.