Lens structure segmentation from AS-OCT images via shape-based learning
Introduction
The crystalline lens is one of the important refractive media in the eyeball, which is a biconvex lens with elasticity. The shape of the lens can be adjusted by the ciliary muscle to focus on nearby or far objects. With age, the nucleus of the lens progressively condenses, expands, and loses its elasticity. Therefore, the adjusting ability of the lens becomes poor, and presbyopia occurs [1]. In addition to presbyopia, cataract is also caused by abnormalities in the lens. Cataract is the largest contributor to global blindness in adults aged 50 years and older in 2020 and is approximately 45% of the cases of global blindness [2]. The primary reason for cataracts is lens opacity caused by the degeneration of lens proteins, which prevents light from being projected onto the retina, resulting in blurred vision. In clinical practice, a standard indicator of cataract grading is the Lens Opacities Classification System III (LOCS III) [3]. However, LOCS III is relatively subjective and the outcomes are affected by the experience of the ophthalmologists. Recently, many studies have focused on the correlation between LOCS III and quantitative nuclear densities calculated from anterior segment optical coherence tomography (AS-OCT) [4,5]. Moreover, nuclear density is proposed to be an objective metric for cataract grading [6,7].
Accurate nucleus segmentation is very important to calculate the nuclear density from AS-OCT. Meanwhile, the segmentations of the nucleus and cortex are also meaningful for the automatic diagnosis of presbyopia and other diseases. However, it is challenging to segment the accurate nucleus because the boundaries between the cortex and the nucleus are fuzzy (see Fig. 1(B)). Zhang et al. [8] proposed a guide-based M-shape convolutional network (G-MNet) to segment the capsule, cortex, and nucleus from AS-OCT images. The G-MNet could segment the high-resolution images guided by the multi-scale and low-resolution segmentation outputs. Cao et al. [9] first extracted the lens area and then segmented the nuclear structure by using a ShuffleSeg network. Finally, they adopted a curve fitting processing to improve the nucleus segmentation. Although these methods can produce nucleus segmentation, low accuracy results are still produced due to the low contrast between the target and background in the image and the lack of object shape constraints in the methods. With the help of the shape information, the nucleus segmentation results can be improved. Thus, Yin et al. [10] proposed a two-step framework by using a deep learning network followed by a post-processing step with a shape template. However, due to individual differences, the segmentation results obtained by the shape template method could not fit the object boundary of each sample well.
The level set is a non-parametric shape representation. In the segmentation task, the segmented boundary can be represented as the zero-level set of the high-dimensional surface corresponding to the level set function [11]. Recently, researchers focus on combining the level set method and deep learning frameworks. The common combination method is considering the output of the deep frameworks as an initial shape, and it can be evolved by level set function to finetune the segmentation result. In addition, the level set methods can be considered recurrent processing, so some algorithms [12], [13], [14] incorporated the level set with the recurrent neural network. The first combination is not an end-to-end solution, and the second combination makes the model complex.
In this paper, we propose an algorithm to incorporate the shape priors with the deep learning network and design an end-to-end framework. A level set function corresponding to the object shape is learned, so the pixel classification strategy commonly used in image segmentation is replaced by a regression strategy, which is similar to object segmentation methods by predicting the signed distance map [15,16]. The main contributions of our algorithm are summarized as follows:
- (1)
A loss function considering shape prior is developed to supervise the deep neural network to learn a level set function, which allows for the end-to-end training and inference processes.
- (2)
A convexity constraint term is designed in the shape-based loss function to ensure the boundaries of the nucleus and cortex segmentation results are convex, which accords with the physiology.
- (3)
A level set normalization method is proposed to alleviate the problem of the weak-contrast boundary between the nucleus and the cortex.
Section snippets
Shape-based learning framework
In the proposed deep learning framework, a level set function is first learned by a U-Net, whose up-sampling consists of Bilinear interpolation followed by a 3 × 3 convolutional layer, as shown in Fig. 2. Notably, the network's last layer is Sigmoid activation, and the output is transformed following output = (sigmoid − 0.5) × 2 to [-1, 1]. This operation is experimentally selected because the results show that the average mean Euclidean Distance (MED) of the models obtained is reduced
Experiments and results
In this section, we present data, implementation details, evaluation metrics, experimental designs, and corresponding results. In the experimental designs, we consider four aspects, including method capability and training stability verification, parameter setting, comparison with state-of-the-art (SOTA) methods, and ablation study. In the experiments of method capability and training stability verification, we obtain an upper bound on the capability of our method. Meanwhile, we verify the
Discussion
Since there are hyperparameters to be selected in deep learning methods while our dataset is not large, we divide the data into DP#2 (containing the training and validation sets) intending to determine these hyperparameters and obtain an upper bound on the method's capabilities. We then retrain and test our proposed model with the selected hyperparameters on data DP#1 and 3 (both containing training and test sets) to verify its stable performance and well generalization to different data
Conclusion
In this paper, we utilize a U-Net to learn a level set function and propose a simple and effective shape-based loss. Our novel loss considers the shape prior of the segmented objects and can naturally fulfill convexity constraints, which improves the segmentation performance. Moreover, we propose a level set normalization method so that the model can focus on the weak-contrast boundary of the objects. Extensive experiments on nucleus and cortex segmentation from AS-OCT images demonstrate the
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (30)
- et al.
The mechanism of presbyopia[J]
Prog. Retin. Eye Res.
(2005) - et al.
New objective lens density quantification method using swept-source optical coherence tomography technology: comparison with existing methods
J. Cataract Refract. Surg.
(2017) - et al.
Causes of blindness and vision impairment in 2020 and trends over 30 years, and prevalence of avoidable blindness in relation to VISION 2020: the right to sight: an analysis for the global burden of disease study[J]
Lancet Glob. Health
(2021) The lens opacities classification system III
Arch. Ophthalmol.
(1993)- et al.
Quantitative analysis of lens nuclear density using optical coherence tomography (OCT) with a liquid optics interface: correlation between OCT images and LOCS III grading
J. Ophthalmol.
(2016) Objective quantification of lens nuclear opacities using swept-source anterior segment optical coherence tomography
Br. J. Ophthalmol.
(2021)Quantitative assessment of lens opacities with anterior segment optical coherence tomography
Br. J. Ophthalmol.
(2009)Guided m-net for high-resolution biomedical image segmentation with weak boundaries
An efficient lens structures segmentation method on AS-OCT images
Automatic segmentation of cortex and nucleus in anterior segment OCT images
Computational Pathology and Ophthalmic Medical Image Analysis
(2018)
Reformulating level sets as deep recurrent neural network approach to semantic segmentation
IEEE Trans. Image Process.
Deep multiphase level set for scene parsing
IEEE Trans. Image Process.
RACE-net: a recurrent neural network for biomedical image segmentation[J]
IEEE J. Biomed. Health Inform.
Topology-preserving shape-based regression of retinal layers in oct image data using convolutional neural networks[C]
Cited by (1)
Asymmetric convolutional multi-level attention network for micro-lens segmentation
2024, Engineering Applications of Artificial Intelligence
- 1
These authors contributed equally to this work.