Development of a Transfer Learning-Based Interactive Tool for Pathological Image Tissue Segmentation

Abstract

Tissue segmentation in pathological images requires high expertise and considerable effort. An interactive semantic segmentation tool was developed that sequentially learns annotation inputs from a user and refines predictions to improve accuracy. The tool utilizes a Random Forest classifier to segment Masson’s Trichrome-stained images into three classes: “Nucleus,” “Collagen,” and “Background.” Experiments to evaluate the effectiveness of the tool were conducted by single trained subject. Evaluation, based on Intersection over Union (IoU) and the number of annotation operations demonstrated that the tool maintained accuracy comparable to manual annotation while improving work efficiency by approximately 1/10. Although transfer learning was attempted to adapt the model to new images, a clear improvement in efficiency was not obtained. A limitation in nucleus recognition accuracy was suggested. Further improvements in both accuracy and efficiency are expected through the addition of shape-based features.