Source-linked AI summary
Deep Learning based Retinal OCT Segmentation
Mike Pekala, Neil Joshi, David E. Freund, Neil M. Bressler, Delia Cabrera DeBuc, Philippe M Burlina
TL;DR
The paper asks whether deep learning can achieve accurate fine-grained automatic segmentation of retinal OCT images. It combines FCN-based layer classification with Gaussian-process regression and evaluates the approach against human annotations and five established algorithms. The proposed methods achieve the smallest reported mean unsigned errors and performance comparable with human annotation on mild retinopathy data.
Problem
The study evaluates whether deep learning can provide effective fine-grained automatic segmentation of retinal OCT images, where retinal-layer boundaries matter for analyzing structural disease manifestations.
Method
The approach combines DenseNet-based fully convolutional networks for per-pixel layer classification with Gaussian-process regression for smooth retinal-surface estimation.
Results
The proposed methods produce the smallest mean unsigned errors and associated standard deviations among the compared methods, with performance comparable to human annotation.
Takeaways & Limitations
FCN semantic segmentation coupled with regression-based post-processing can produce OCT segmentation results on par with human capabilities on the evaluated mild-retinopathy dataset.
Takeaways & Limitations
The dataset represents only mild diabetic retinopathy, and strict comparisons are limited because reference algorithms were developed on datasets with potentially different resolutions, noise, and artifacts.
Abstract
from arXiv · showhide
Our objective is to evaluate the efficacy of methods that use deep learning (DL) for the automatic fine-grained segmentation of optical coherence tomography (OCT) images of the retina. OCT images from 10 patients with mild non-proliferative diabetic retinopathy were used from a public (U. of Miami) dataset. For each patient, five images were available: one image of the fovea center, two images of the perifovea, and two images of the parafovea. For each image, two expert graders each manually annotated five retinal surfaces (i.e. boundaries between pairs of retinal layers). The first grader's annotations were used as ground truth and the second grader's annotations to compute inter-operator agreement. The proposed automated approach segments images using fully convolutional networks (FCNs) together with Gaussian process (GP)-based regression as a post-processing step to improve the quality of the estimates. Using 10-fold cross validation, the performance of the algorithms is determined by computing the per-pixel unsigned error (distance) between the automated estimates and the ground truth annotations generated by the first manual grader. We compare the proposed method against five state of the art automatic segmentation techniques. The results show that the proposed methods compare favorably with state of the art techniques, resulting in the smallest mean unsigned error values and associated standard deviations, and performance is comparable with human annotation of retinal layers from OCT when there is only mild retinopathy. The results suggest that semantic segmentation using FCNs, coupled with regression-based post-processing, can effectively solve the OCT segmentation problem on par with human capabilities with mild retinopathy.
1 Introduction
OCT provides high-resolution, non-invasive views of retinal structure relevant to ophthalmic, neurological, and vascular disorders. Against a backdrop of classical OCT segmentation methods and emerging deep-learning approaches, this work proposes FCN–Gaussian-process segmentation and evaluates it against established algorithms and human annotation.
- Motivation: OCT captures micron-scale retinal structure non-invasively, and structural abnormalities are associated with ophthalmic, neurodegenerative, and vascular disorders.The introduction highlights AMD, retinal thinning, and other layer changes as clinically relevant examples.
- Related work: Automated OCT segmentation has largely used classical methods, while recent studies have begun applying convolutional neural networks and hybrid architectures.Prior work includes graph-based methods, ConvNets, cascaded U-Net-like models, and ConvNet–graph combinations.
- Contribution: The study introduces DenseNet-based fully convolutional networks combined with Gaussian process regression for retinal OCT segmentation.The method is evaluated on a publicly available U. of Miami dataset against methods of record and human annotation.
- Contribution: The proposed approach reports the smallest unsigned boundary estimation errors and performance on par with a human annotator on the evaluated dataset.The authors connect this result to potential clinical relevance because retinal-layer manifestations can accompany ophthalmic, neurological, and vascular disorders.
2 Methods
The study uses manually annotated OCT images from patients with mild diabetic retinopathy and estimates retinal surfaces through per-pixel FCN classification followed by smoothness-constrained Gaussian-process regression. Performance is assessed with patient-level cross-validation against human annotations and five established automated algorithms.
- 2.1 Data: The dataset contains 50 OCT images from 10 patients with mild non-proliferative diabetic retinopathy, covering foveal, perifoveal, and parafoveal regions.Each image has five manually annotated retinal surfaces from two expert graders; the first grader supplies ground truth and the second measures inter-operator agreement.
- 2.1 Data: A retinal surface is defined as the boundary between adjacent retinal layers, with five numbered surfaces annotated per image.The study uses surfaces 1, 2, 4, 6, and 11 from the public dataset.
- 2.2 Segmentation approach: The first segmentation stage uses a DenseNet-based FCN to produce per-pixel retinal-layer class estimates with the same spatial dimensions as the input image.DenseNet connections allow layers to process outputs from all preceding layers, propagating features through the network.
- 2.2 Segmentation approach: A Gaussian-process regression stage converts classification outputs into smooth retinal-surface estimates, addressing artifacts that can cause missing or duplicate surfaces.The radial-basis-function kernel used variance and length-scale values of 50 pixels; the combined method is termed SEG+REG.
- 2.3 Reference methods: Five reference algorithms—Spectralis 6.0, IOWA Reference Algorithm, AURA, Dufour’s Bern algorithm, and OCTRIMA3D—are included for comparison.Their annotations may not span the entire OCT image, so evaluation is restricted to regions where all automated methods have valid estimates.
- 2.4 Evaluation methods and metrics: The evaluation uses 10-fold patient-level cross-validation, training on 45 images from nine patients and testing on the remaining patient’s five images in each fold.Accuracy is measured with mean unsigned and signed per-pixel errors against the first grader’s annotations, within regions valid for all automated algorithms.
3 Results
The proposed SEG and SEG+REG methods match human performance in aggregate and compare favorably with established OCT segmentation algorithms. Their strongest reported behavior is on inner retinal surfaces, although signed errors suggest some systematic layer-support overestimation.
- Aggregate performance: The proposed methods match human performance in aggregate and compare favorably with other algorithms of record.Table 2 reports mean unsigned errors by surface, along with average and maximum values across testing data.
- Surface-specific performance: The proposed methods show particularly good performance on inner retinal surfaces.The result is reported from the mean unsigned-error comparisons across surfaces.
- Signed errors: Surface 1 has a relatively large positive signed error, while surface 11 has a relatively large negative signed error.The authors interpret these signed-error patterns as possible slight overestimation of retinal-layer support.
4 Discussion
The proposed FCN-based segmentation with GP regression produced errors comparable to human annotation and favorable relative to methods of record, while comparisons require caution because those methods were developed on differing datasets. The approach is relatively simple and may extend to volumetric OCT and additional structures, but the study covered only mild diabetic retinopathy.
- FCN semantic segmentation coupled with GP regression effectively addressed fine-grained automated OCT segmentation.The approach combines DenseNet FCNs with regression-based post-processing.
- The proposed methods achieved the smallest mean unsigned errors and associated standard deviations, with overall performance comparable to human annotation.The discussion reports favorable comparisons with state-of-the-art methods and human performance.
- Comparisons with algorithms of record require caution because those algorithms were optimized on datasets differing in resolution, noise, and artifacts.
- The proposed approaches are relatively simple and can be extended to direct 3D volumetric analysis and identification of structures such as drusen or lesions.
- The study’s dataset represented only mild diabetic retinopathy, motivating future evaluation on more advanced conditions.The authors also anticipate improvement from additional training data and more exhaustive hyperparameter selection.
5 Conclusion
The paper proposes automated OCT segmentation methods using FCNs with regression-based post-processing. Results suggest performance on par with human capabilities and at least as good as the prior methods considered.
- FCNs coupled with regression-based post-processing produced OCT segmentation results on par with human capabilities and met or exceeded the prior methods considered.