[4Q2-OS-5] 医療におけるAIの社会実装に向けて

[4Q2-OS-5-04] Proposition of semi-supervised sequential image segmentation method using selective annotation

〇Eichi Takaya^1,2,3, Satoshi Kurihara¹ (1. Graduate School of Science and Technology, Keio University, 2. St. Marianna University Graduate School of Medicine, 3. AI Lab, Tohoku University Hospital)

[[Online]]

Keywords:Medical image, Fractal Analysis, Semi-supervised learning, Deep learning

Annotation of medical images such as MRI and CT images is essential for measuring the effectiveness of cancer treatment and identifying target areas for radiotherapy.
In recent years, there have been many cases in which annotation is required in the pre-processing of image data in the field of medical AI research.
However, medical professionals have a huge amount of work to do in their daily clinical practice, and the amount of time they can devote to annotation work is limited.
In this study, we propose a method to realize the automatic segmentation of sequential images with minimal annotation.
Specifically, the fractal dimension of a sequence of images is calculated, and a few consecutive slices with large values are selected for annotation. A semi-supervised segmentation method is applied to the remaining slices.
Experiments on left atrial segmentation using a public cardiac MRI dataset demonstrate the effectiveness of the proposed method.