Year: 2026
Type: dissertation
Abstract: Pathological diagnosis is the act of diagnosing diseases by observing tissues and cells collected from patients under a microscope. Pathological diagnosis requires a high level of expertise and extensive experience; however, there is a shortage of such pathologists. In particular, there is a severe shortage of pathologists specializing in rare diseases, which is a global issue. Therefore, there is a social need for the development of artificial intelligent (AI)-based pathological diagnosis support systems. In this study, we evaluated the feasibility of a pathological diagnosis support system by verifying an AI-based pathological diagnosis algorithm targeting rare diseases. The target diseases were rare muscle diseases and malignant lymphoma. The study used digitized pathological image data stored at hospitals. Muscle pathology images and lymphoma images were obtained from the National Center of Neurology and Psychiatry and Kurume University, respectively. We attempted to classify hereditary and non-hereditary muscle diseases using AI. For lymphoma, we aimed to differentiate between diffuse, follicular and benign subtypes of malignant lymphoma. In this study, we used a deep convolutional neural network (CNN), which is one of the useful deep learning model architectures. Evaluation method used was a 5-fold cross validation for both muscle diseases and malignant lymphoma. Evaluation metrics included the area under the receiver operating characteristic curve (AUC) and accuracy. Additionally, as a benchmark, the results of disease classification performed by pathologists based on images were compared with the accuracy of the deep learning model, thereby demonstrating the usefulness of the proposed models. As a result, the model for distinguishing between hereditary and non-hereditary diseases in muscle pathology achieved an AUC of 0.996, while an average AUC of 0.950 was achieved for distinguishing between diffuse, follicular, and benign lymphomas. The proposed model obtained the accuracy of 96.9% in classifying muscle diseases, in contrast to the highest accuracy of 93.8% among specialists. On the other hand, the accuracy of the deep learning model was 97.0%, while the highest accuracy among hematopathologists was 83.3% for malignant lymphoma. In both diseases, the deep CNN model achieved the accuracies comparable to or higher than those of specialists. These results suggest that the deep CNN models proposed in this study have a potential for practical application in clinical practice.
Author: Kabeya, Yoshinori
Institution: Shiga University
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162720081
