Formula-Driven Data Augmentation and Partial Retinal Layer Copying for Retinal Layer Segmentation

Tsubasa Konno¹, Takahiro Ninomiya², Kanta Miura¹, Koichi Ito¹, Noriko Himori², Parmanand Sharma², Toru Nakazawa², Takafumi Aoki¹

¹Graduate School of Information Sciences, Tohoku University
²Department of Ophthalmology, Graduate School of Medicine, Tohoku University
The 11th OMIA Workshop on MICCAI 2024
^*Indicates Equal Contribution

Abstract

Major retinal layer segmentation methods from OCT images assume that the retina is flattened in advance, and thus cannot always deal with retinas that have changes in retinal structure due to ophthalmopathy and/or curvature due to myopia. To eliminate the use of flattening in retinal layer segmentation for practicality of such methods, we propose novel data augmentation methods for OCT images. Formula-driven data augmentation (FDDA) emulates a variety of retinal structures by vertically shifting each column of the OCT images according to a given mathematical formula. We also propose partial retinal layer copying (PRLC) that copies a part of the retinal layers and pastes it into a region outside the retinal layers. Through experiments using the OCT MS and Healthy Control dataset and the Duke Cyst DME dataset, we demonstrate that the use of FDDA and PRLC makes it possible to detect the boundaries of retinal layers without flattening even retinal layer segmentation methods that assume flattening of the retina.

Formula-Driven Data Augmentation (FDDA)

Formula-Driven Data Augmentation (FDDA) is a data augmentation method that emulates a variety of retinal structures based on mathematical formulas, and increase the variability of retinal shapes in training data. FDDA changes the the position, the tilt, and the curvature of the retina by vertically shifting each column of the OCT images according to a given mathematical formula. Specifically, FDDA shifts each column of an OCT image according to a simple combination of $N$th-order functions. In FDDA, the image center is the origin of $N$th-order functions, and the amount of pixel shift is determined based on these functions. Since this process only shifts the pixels by the amount of shift determined by simple functions, the labels of the OCT image after data augmentation can be easily obtained by shifting the boundary labels in the same way.

Fig. 1. Overview of FDDA using only the first-order shift.

Fig. 2. Examples of applying FDDA to OCT images, where each of the zeroorder, first-order, and second-order shifts is applied to the input image for simplicity: (a) the zero-order shift, (b) the first-order shift, (c) the second-order shift, and (d) the combined shift, and an example of applying RandomAffine for comparison. Colored lines on each image indicate the annotated boundaries between the retinal layers.

Partial Retinal Layer Copying (PRLC)

Partial Retinal Layer Copying (PRLC) is a data augmentation method that reproduces the background noise of OCT images, and reduces false detection in the background region. Specifically, PRLC copies a part of the retinal layers and pastes it into a region outside the retinal layers. The parameters in PRLC are the number of retinal layers $l$ to be copied and the width $W$ of the retinal layers. In this paper, $l$ is set to 1 to 3 and $W$ is set to 20 to $N_2$, which are selected at random from the corresponding ranges. First, the target retinal layer is randomly determined and the adjacent retinal layers are selected according to $l$. Next, the retinal layer region to be copied is determined according to $W$. Then, we paste the above retinal layer region at a random position in the background region where no retinal layer labels are assigned. If there is no space in the background region to paste the retinal layer region, we repeat the process from the first step.

Fig. 3. Examples of applying PRLC to OCT images, where the red dashed box indicates the pasted retinal layer area. An example of applying CutMix is also shown for comparison. Colored lines on each image indicate the annotated boundaries between the retinal layers.

Evaluation

We demonstrate the effectiveness of the proposed method for retinal layer segmentation by applying FDDA and PRLC to conventional methods. We use the OCT MS and Healthy Control (MSHC) dataset and the Duke Cyst DME (Duke DME) dataset in the experiments. The accuracy of each method is evaluated by the mean absolute distance (MAD) between the detected boundary and the ground truth.

Table 1. Experimental results of each method for MSHC dataset and Duke DME dataset. The units for MAD are $\mu$m.

BibTeX

@article{Konno-OMIA-2024, author = "Konno, T. and Ninomiya, T. and Miura, K. and Ito, K. and Himori, N. and Sharma, P. and Nakazawa, T. and Aoki, T.", title = "Formula-Driven Data Augmentation and Partial Retinal Layer Copying for Retinal Layer Segmentation", journal = "Proc. Ophthalmic Medical Image Analysis Workshop on MICCAI", year = "2024", month = oct }