Presentation Information
[A-15-09]Mask-Supervised Attention Learning for Lesion Region Analysis in RDI Endoscopic Images
◎△HAOWEN BAI1, Suzuki Hiroyuki2, Nabilah Hannani3, Jiang Pei3, Itoi Yuki4, Aung Paing Moe4, Takeuchi Yoji4, Kuribayashi Shiko4, Obi Takashi3, Uraoka Toshio4 (1. Graduate School of Informatics, Gunma Univ, 2. Center for Math. & Data Sci., Gunma Univ, 3. Institute of Integrated Research, Institute of Science Tokyo, 4. Graduate School of Medicine, Gunma Univ)
Keywords:
Endoscopic Image Analysis,Deep Learning,Ulcerative Colitis,Red Dichromatic Imaging,Attention-guided Learning,Pretrained Model,Image Classification
In recent years, deep learning-based computer-aided diagnosis systems for endoscopic imaging have been extensively studied. Red Dichromatic Imaging (RDI) endoscopy is an image enhancement technology capable of highlighting blood vessels and inflamed areas, holding great promise for application in the diagnostic support of ulcerative colitis (UC) . However, the limited availability of RDI training data poses a significant challenge in constructing highly accurate classification models. Furthermore, the reliability of the diagnostic rationale is often compromised when the model attends to non-lesion areas. Aiming to develop an artificial intelligence (AI) system for grading UC from RDI endoscopic images, this study proposes a novel RDI image classification method that integrates transfer learning—based on a model pre-trained on a public image dataset with attention-guided learning utilizing lesion region masks.