Retinal Layer Segmentation in OCT Images With 2.5D Cross-slice Feature Fusion Module for Glaucoma Assessment
Authors: Hyunwoo Kim, Heesuk Kim, Wungrak Choi, Jae-Sang Hyun
Organizations: Department of Mechanical Engineering, Yonsei University, Seoul, 03722, South Korea · Institute of Vision Research, Department of Ophthalmology, Severance Hospital, Yonsei University, Seoul, 03722, South Korea
For accurate glaucoma diagnosis and monitoring, reliable retinal layer segmentation in OCT images is essential. However, existing 2D segmentation methods often suffer from slice-to-slice inconsistencies due to the lack of contextual information across adjacent B-scans. 3D segmentation methods are better for capturing slice-to-slice context, but they require expensive computational resources. To address these limitations, we propose a 2.5D segmentation framework that incorporates a novel cross-slice feature fusion (CFF) module into a U-Net-like architecture. The CFF module fuses inter-slice features to effectively capture contextual information, enabling consistent boundary detection across slices and improved robustness in noisy regions. The framework was validated on both a clinical dataset and the publicly available DUKE DME dataset. Compared to other segmentation methods without the CFF module, the proposed method achieved an 8.56% reduction in mean absolute distance and a 13.92% reduction in root mean square error, demonstrating improved segmentation accuracy and robustness. Overall, the proposed 2.5D framework balances contextual awareness and computational efficiency, enabling anatomically reliable retinal layer delineation for automated glaucoma evaluation and potential clinical applications.
Figures & tables
Figure 1: Overview of the proposed method.
Figure 2: Pre-processing steps for OCT images.
Figure 3: Overview of cross-slice feature fusion (CFF) module with input F and output Ffused .
Normal
Glaucoma
Other disease
Total
Train dataset
4
12
4
20
Validation dataset
0
2
0
2
Test dataset
2
6
2
10
Total
6
20
6
32
Table 1: Distribution of dataset cases across training, validation, and test sets.
Figure 4: Example of our dataset. (a) Structure of the macular cube. (b) B-scan from the middle of the macular cube with 5 target retinal layer boundaries annotated by experienced clinicians.
(a) MAD (Std)
Method
ILM
RNFL-GCL
IPL-INL
OPL-ONL
ONL-IS
Average
Graph [ 2 ]
3.21 (1.57)
9.06 (3.04)
5.74 (2.77)
4.69 (3.35)
2.26 (0.86)
4.99 (1.87)
Basic UNet [ 28 ]
2.43 (0.80)
3.26 (1.62)
2.84 (1.00)
2.80 (0.54)
2.03 (0.55)
2.67 (0.80)
ReLayNet [ 30 ]
2.53 (0.91)
3.09 (0.89)
2.89 (0.97)
2.85 (0.59)
2.06 (0.63)
2.69 (0.68)
FCRN [ 9 ]
3.07 (0.90)
2.66 (0.99)
2.86 (0.46)
2.98 (0.45)
2.05 (0.31)
2.64 (0.58)
Ours
2.36 (0.80)
2.93 (0.67)
2.56 (0.39)
2.70 (0.47)
1.85 (0.29)
2.48 (0.48)
Table 2: Evaluation results on our dataset among multiple methods. Mean absolute distance (MAD) and root mean squared error (RMSE) of the boundary positions are measured in pixel units, with mean and standard deviation. Lower is better. The bolded values are the best performing values.
Figure 5: Qualitative results of the comparison methods on our dataset. The rows represent three consecutive B-scans. An enlarged region is shown in the upper-right corner of each image.
(a) MAD (Std)
Method
ILM
RNFL-GCL
IPL-INL
OPL-ONL
ONL-IS
Average
Graph [ 2 ]
1.86 (0.40)
8.05 (2.96)
13.37 (5.05)
16.37 (6.52)
3.05 (1.00)
8.54 (3.07)
Basic UNet [ 28 ]
4.44 (4.07)
5.66 (2.44)
7.93 (3.28)
12.83 (1.46)
7.18 (1.46)
7.61 (3.15)
ReLayNet [ 30 ]
3.90 (1.45)
10.51 (4.99)
10.24 (4.66)
9.26 (3.38)
14.98 (7.62)
9.78 (4.01)
FCRN [ 9 ]
3.70 (2.18)
6.72 (2.81)
5.30 (1.71)
5.66 (1.68)
4.56 (1.35)
5.19 (1.76)
Ours
2.96 (1.10)
5.65 (1.25)
5.04 (1.39)
5.10 (1.78)
4.67 (1.28)
4.68 (1.11)
Table 3: Evaluation results on the DUKE DME dataset among multiple methods. Mean absolute distance (MAD) and root mean squared error (RMSE) of the boundary positions are measured in pixel units, with mean and standard deviation. Lower is better. The bolded values are the best performing values.
Figure 6: Qualitative results of the comparison methods on the DUKE DME dataset. The cyan regions in the ground truth indicate the fluid regions.
Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, China · College of Information Engineering, Zhejiang University of Technology, Hangzhou, China
Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China · Shenzhen Loop Area Institute, Shenzhen, China · Jiaxing Research Institute, Southern University of Science and Technology, Jiaxing, China