A comparison of deep learning U-Net architectures for posterior segment OCT retinal layer segmentation

Abstract

Deep learning methods have enabled a fast, accurate and automated approach for retinal layer segmentation in posterior segment OCT images. Due to the success of semantic segmentation methods adopting the U-Net, a wide range of variants and improvements have been developed and applied to OCT segmentation. Unfortunately, the relative performance of these methods is difficult to ascertain for OCT retinal layer segmentation due to a lack of comprehensive comparative studies, and a lack of proper matching between networks in previous comparisons, as well as the use of different OCT datasets between studies. In this paper, a detailed and unbiased comparison is performed between eight U-Net architecture variants across four different OCT datasets from a range of different populations, ocular pathologies, acquisition parameters, instruments and segmentation tasks. The U-Net architecture variants evaluated include some which have not been previously explored for OCT segmentation. Using the Dice coefficient to evaluate segmentation performance, minimal differences were noted between most of the tested architectures across the four datasets. Using an extra convolutional layer per pooling block gave a small improvement in segmentation performance for all architectures across all four datasets. This finding highlights the importance of careful architecture comparison (e.g. ensuring networks are matched using an equivalent number of layers) to obtain a true and unbiased performance assessment of fully semantic models. Overall, this study demonstrates that the vanilla U-Net is sufficient for OCT retinal layer segmentation and that state-of-the-art methods and other architectural changes are potentially unnecessary for this particular task, especially given the associated increased complexity and slower speed for the marginal performance gains observed. Given the U-Net model and its variants represent one of the most commonly applied image segmentation methods, the consistent findings across several datasets here are likely to translate to many other OCT datasets and studies. This will provide significant value by saving time and cost in experimentation and model development as well as reduced inference time in practice by selecting simpler models.

Automatic segmentation and quantification of the optic nerve on MRI using a 3D U-Net

“…Despite the basic nature of our method, it demonstrated high performance and satisfactory segmentations. While more advanced variants of the U-Net could be further explored, previous research has indicated that they may not necessarily result in significant performance improvements and meanwhile increase complexity and require a higher demand on computational resources 43 …”

“…While more advanced variants of the U-Net could be further explored, previous research has indicated that they may not necessarily result in significant performance improvements and meanwhile increase complexity and require a higher demand on computational resources. 43 The ON quantification method developed in this study was validated based on manual reference measurements. However, a direct comparison with manual measurements is not straightforward as the offset between the model's origin and manual reference varies for each nerve, and manual measurements are subjective and sensitive to errors.…”

Characterisation of the normal human ganglion cell–inner plexiform layer using widefield optical coherence tomography

“…A segmentation pipeline using a deep learning approach was applied to segment the two retinal layer boundaries delineating the GCIPL, the retinal nerve fibre layer–ganglion cell layer (RNFL–GCL) and inner plexiform layer–inner nuclear layer (IPL–INL) boundaries (Figure 1). In brief, this consisted of cropped OCT images, used to improve generalisability to unseen OCT data, input into a typical encoder–decoder neural network (U‐Net) 37,38 . The output of this process was the per‐pixel classification of OCT B‐scan images into one of three classes, representing the regions separated by the RNFL–GCL and IPL–INL boundaries, which was subsequently used to infer boundary locations via graph search.…”

“…In brief, this consisted of cropped OCT images, used to improve generalisability to unseen OCT data, input into a typical encoder-decoder neural network (U-Net). 37,38 The output of this process was the per-pixel classification of OCT B-scan images into one of three classes, representing the regions separated by the RNFL-GCL and IPL-INL boundaries, which was subsequently used to infer boundary locations via graph search. This process was first applied with the raw widefield OCT image to produce initial predicted layer boundary positions, which were subsequently refined by repeating the process using a flattened OCT image, to normalise the appearance of the retinal tissue structure in the image provided to the network.…”

“…This process was first applied with the raw widefield OCT image to produce initial predicted layer boundary positions, which were subsequently refined by repeating the process using a flattened OCT image, to normalise the appearance of the retinal tissue structure in the image provided to the network. 37 This was performed by fitting a second-order polynomial to the IPL-INL boundary, and A-scans were translated using this polynomial to flatten the OCT image prior to re-input to the network and a subsequent graph search. The new boundary positions were then transformed to their original coordinates by undoing the flattening to give the segmented RNFL-GCL and IPL-INL boundaries.…”