Deep learning for the segmentation of preserved photoreceptors on en face optical coherence tomography in two inherited retinal diseases

: The objective quantification of photoreceptor loss in inherited retinal degenerations (IRD) is essential for measuring disease progression, and is now especially important with the growing number of clinical trials. Optical coherence tomography (OCT) is a non-invasive imaging technology widely used to recognize and quantify such anomalies. Here, we implement a versatile method based on a convolutional neural network to segment the regions of preserved photoreceptors in two different IRDs (choroideremia and retinitis pigmentosa) from OCT images. An excellent segmentation accuracy (~90%) was achieved for both IRDs. Due to the flexibility of this technique, it has potential to be extended to additional IRDs in the future.

12. G. Staurenghi, S. Sadda, U. Chakravarthy, and R. F. Spaide, "Proposed lexicon for anatomic landmarks in normal posterior segment spectral-domain optical coherence tomography," Ophthalmology 121(8), 1572-1578 (2014). 13. D. G. Birch, Y. Wen, K. Locke, and D. C. Hood, "Rod sensitivity, cone sensitivity, and photoreceptor layer thickness in retinal degenerative diseases," Invest. Ophthalmol. Vis. Sci. 52(10), 7141-7147 (2011). 14. G. Liu, H. Li, X. Liu, D. Xu, and F. Wang, "Structural analysis of retinal photoreceptor ellipsoid zone and postreceptor retinal layer associated with visual acuity in patients with retinitis pigmentosa by ganglion cell analysis combined with OCT imaging," Medicine (Baltimore) 95 ( Huang, and Y. Jia, "Advanced image processing for optical coherence tomographic angiography of macular diseases," Biomed. Opt. Express 6(12), 4661-4675 (2015 [2,8,11] are used for assessment of disease progression in the clinical practice. Since OCT is the only one to provide depth-resolved information of retinal tissue, it is the most solid existing technology for imaging and quantification of photoreceptor preservation. In OCT images, the second hyper-reflective layer of the outer retina, identified as the ellipsoid zone (EZ) of the photoreceptors [12], is the structure most suitable to assess photoreceptor damage [13]. Numerous image processing techniques have been reported in recent years to detect and quantify the extent of EZ damage in IRDs [14], macular telangiectasia [15] and ocular trauma [16]. Previously, we have developed en face methods that use OCT images to detect EZ loss in mild diabetic retinopathy by fuzzy logic [17] and choroideremia by a random forest classifier [18]. However, the pattern of photoreceptor integrity can present differently in each retinal pathology. For example, with retinitis pigmentosa the best strategy is to detect the preserved EZ boundary since the degeneration starts in the mid-periphery and constricts centrally to leave a round-shaped "island" of preserved EZ centered at the fovea [2]. For other diseases, such as Stargardt Dystrophy, where photoreceptor atrophy starts centrally, it is more feasible to detect EZ loss. The pattern of EZ atrophy can present with complex shapes such as with choroideremia, which shows initial loss in the periphery of the macula, scalloped edges [19] and outer retinal tubulations [20]. Consequently, image processing methods developed targeting a certain disease assuming certain ad hoc rules are not generalizable and typically do not perform as well for patients with a different IRD.
With the purpose of developing a single method that is adaptable to different retinal conditions, we have implemented a deep learning platform that can be trained for more than one IRD (herein, retinitis pigmentosa and choroideremia) to detect the areas of preserved EZ. Our approach uses a segmentation method consisting of sliding-window binary classification of OCT B-scan sections by a convolutional neural network (CNN). In the context of deep learning, the segmentation problem is that of finding the pixels that belong to a certain semantic class that the network has been trained to recognize (e.g. defect tissue vs. healthy tissue). Here, we use a CNN trained from B-scan patches enclosing sections of the EZ, each of which is labeled based on the appearance of en face images at the patch's central A-line position. Further bimodal thresholding of probability maps by an Otsu scheme and morphological operations provided binary maps of the segmented preserved photoreceptor areas with high accuracy compared to manual segmentation by an expert grader.

Study population
Twenty subjects diagnosed with chorideremia and twenty-two diagnosed with retinitis pigmentosa were recruited from the Ophthalmic Genetics clinic at the Casey Eye Institute at the Oregon Health & Science University (OHSU). The protocol was approved by the Institutional Review Board/Ethics Committee of OHSU and the research adhered to the tenants of the Declaration of Helsinki.

Data acquisition
Macular scans covering a 6 mm × 6 mm area were acquired by a 70-kHz, 840-nm-wavelength spectral-domain OCT system (Avanti RTVue-XR, Optovue Inc.) within 2.9 seconds. The AngioVue version 2016.2.0.35 software was used to acquire optical coherence tomography angiography (OCTA) scans. In the fast transverse scanning direction, 304 A-scans were sampled to form a B-scan and two repeated B-scans were acquired at each lateral location. A total of 304 locations were scanned in the slow transverse direction to form a 3D data cube. Axial resolution in AngioVue is 5 µm but digital pixel sampling is 3 µm. Structural OCT data was obtained generated by order to remo sets of volum merged by mo

Image pr
The algorithm training of a step uses the This interface segmented [2 retina were re from B-scans and morpholo EZ in the pos processes in d 2017a (M (http://www.v

Pre-proc
The inner lim Bruch's-mem directional gr Bruch's mem section includ membrane an (interdigitatio Since the identified by absorption of [ Fig. 2

En face p
The generic activation and The convolut kernels to ext receptive field neurons), con weight across generated, co compress the addition to th introduce the times, a fully of classificatio Here, we [28] to classif A-line. This n (one max poo soft-max clas trained on Bhand. The net validation set a weight dec reduced to 0.0 1 × 2 vector c or EZ loss). T and a NVIDIA Ten eyes w network for e xtraction hes of 33 × 3 Fig. 2(C). Ea o the label assi to the CNN in onding 33 × 33 t or last A-line Z loss patches w 5. Patch extraction ated in Fig. 2

Results
The CNN w pigmentosa an completed the The Jacca automatically was used to a large values m was observed maps due to were properly to manual gr pigmentosa an