Automatic Generation of AI-based Cancer Pathology Data and Highly Accurate Colorectal Cancer Pathology Diagnosis Support

Abstract

Owing to the increase in the number of pathological diagnoses and the shortage of pathologists, the burden on pathologists has been increasing. Accordingly, support systems are expected to be utilized for analyzing pathological images using deep learning to reduce the burden on pathologists. However, the deep learning model needs to be trained using a dataset consisting of a large number of cases, to improve its performance. However, the creation of such a dataset is labor-intensive and time-consuming. Thus, an efficient method for creating datasets is required to create large datasets for future practical use. In this paper, we propose a method for creating datasets using image segmentation based on deep learning. First, we investigated whether the discriminative performance of the deep learning model can be improved by using a narrow-band light source for photographing pathological specimens. Consequently, the correct response rate was 0.93 when a white LED was used as the light source and the image was used as the input, and 0.95 when two narrow-band light sources of wavelengths 500 nm and 570 nm were used as the light sources and the image was used as the input. This indicates that using a specific narrow-band light source can improve the discrimination performance of the deep learning model compared with using white LEDs as the light source. In addition, we efficiently constructed a large and precise dataset consisting of 1018 colorectal pathology images (2028 images) and pixel-by-pixel annotation information using a dataset creation method based on image segmentation via deep learning. In contrast to the conventional handwritten annotation process, which takes an average of 520 s, the proposed method takes an average of 137 s; thus, the creation of the database is accelerated. We trained a deep learning model using the dataset of colorectal pathology specimen images created in this study. The deep learning model was trained to classify images obtained by segmenting the large-sized pathological specimen images into those containing malignant tumors and those not containing malignant tumors. The results of the diagnostic accuracy of the model were as follows: a sensitivity of 95.2%, a specificity of 97.1%, a positive predictive value of 95.29%, and a negative predictive value of 97.06%. The percentage of correct classification was 0.97, and the AUC was 0.99. In this study, we constructed a pathological specimen imaging system using narrow-band light sources of two specific wavelengths as the imaging light source, and created a dataset with high accuracy semi-automatically via image segmentation using deep learning. In addition, we constructed a system that can efficiently and semi-automatically create a large and precise dataset consisting of colorectal pathology images and pixel-by-pixel annotation information. We evaluated these systems and confirmed that they could classify colorectal pathology specimen images as accurately and quickly as or more accurately than pathologists. Thus, we demonstrated their usefulness as a pathology image analysis support system.

Keywords: Cancer Pathology, Diagnosis, Artificial intelligence, Deep learning, Human-machine interface, Digital transformation

DOI: 10.54941/ahfe1001787