東北大学 大学院情報科学研究科 情報基礎科学専攻 計算機構論分野
(東北大学 工学部 電気情報物理工学科 情報工学コース)
青木・伊藤(康)研究室

A Study of Data Augmentation for Retinal Layer Segmentation from Retinal OCT Images

Tsubasa Konno (Tohoku University) , Takahiro Ninomiya (Tohoku University) , Kanta Miura (Tohoku University) , Koichi Ito (Tohoku University) , Noriko Himori (Tohoku University) , Parmanand Sharma (Tohoku University) , Toru Nakazawa (Tohoku University) , Takafumi Aoki (Tohoku University)
電気関係学会東北支部連合大会, August 2024.
Abstract

In ophthalmologic examinations, optical coherence tomography (OCT) images are used to measure thickness of retinal layers that is a significant criterion to diagnose ophthalmopathy. Since itis difficult even for experts to manually measure the thickness from OCT images, retinal layer segmentation methods [1, 2] have been investigated to detect retinal layers from OCT images. Due to lack of labeled training data, these methods cannot deal with a variety of retinal shapes. Furthermore, these methods may falsely segment background noise as retinal layers. To address the above problems, in this paper, we pro- pose two data augmentation methods: formula-driven data augmentation (FDDA) and partial retinal layer copying (PRLC). FDDA emulates a variety of retinal shapes based on mathematical formulas, and increases the variability of retinal shapes in the training data. PRLC duplicates a part of retinal layers and overlays it on the background region as noise, and reduces false segmentation in the background region. Through a set of experiments using public datasets, we demonstrate the effectiveness of the proposed methods.

戻る