機械知覚&ロボティクスグループ
中部大学

Deep Learning 学術論文(E)

Detection of Double Layer Sign in OCT by Semi-supervised semantic segmentation

Author
Yuka Kihara, Mengxi Shen, Yingying Shi, Xiaoshuang Jiang, Liang Wang, Rita Laiginhas, Cancan Lyu, Jin Yang, Jeremy Liu, Rosalyn Morin, Hironobu Fujiyoshi, Giovanni Gregori, Philip J Rosenfeld, Aaron Y Lee
Publication
Investigative Ophthalmology & Visual Science, Vol.64, 3364, 2023

Download: PDF (English)

Purpose : The double layer sign (DLS) on structural B-scans is an important feature of type 1 macular neovascularization (MNV) in age-related macular degeneration (AMD), and annotating these DLSs is both time consuming and requires expertise. We applied a semi-supervised method named Cross Pseudo Supervision (CPS) to leverage both labeled and unlabeled data for training.
Methods : CPS imposes consistency on two segmentation networks that share the same structure but are initialized differently. We built the model using swept-source optical coherence tomography (SS-OCT) and spectral-domain (SD)-OCT scans, separately using eyes with and without type 1 nonexudative MNV (neMNV) that was confirmed on SS-OCT angiography (SS-OCTA) imaging. For the SS-OCT scans, eyes were imaged using SS-OCTA (PLEX Elite 9000, Carl Zeiss Meditec Inc, Dublin, CA) 6x6mm scans. The scans consisted of 500 A-scans per B-scan with each B-scan repeated twice at each B-scan position along the y-axis. For the SD-OCT scans, eyes were imaged using the 200X200 macular cube protocol (Cirrus HD-OCT, Carl Zeiss Meditec Inc) resulting in a dataset with uniformly spaced A-scans organized as 200 A-scans in each B-scan with 200 horizontal B-scans along the y-axis. For the SS-OCT dataset, drusen annotations were also provided.
Results : A total of 251 eyes (211 patients) and 126 eyes (126 patients) were included in the SS-OCT and SD-OCT datasets, respectively. Dataset details and performance are summarized in the figure. The improvements of our method over the supervised method are 3.55% and 2.03% under 1/64 and 1/8 partition protocols in the SS-OCT dataset, and 3.44% and 2.77% in the SD-OCT dataset for DLS class. When all the annotated images are used as labeled data, the improvement in SS-OCT was 1.93%, which slightly outperforms our previous work with Vision Transformer, and the improvement was 0.67% in SD-OCT dataset.
Conclusions : We conducted experiments using Cross Pseudo Supervision to leverage both labeled and unlabeled data for learning. The experiments show the effectiveness of the semi-supervised approach on the task of detecting DLSs.

前の研究 次の研究