Computerized macular pathology diagnosis in spectral domain optical coherence tomography scans based on multiscale texture and shape features

摘要: Spectral-domain optical coherence tomography (SD-OCT) is a noncontact, noninvasive, three-dimensional (3D) imaging technique that performs sectioning at micrometer resolution. It widely used in ophthalmology for identifying the presence of disease and its progression.1 This technology measures back-scattering tissues, making it possible to visualize intraocular structures diagnose ocular diseases, such as glaucoma macular hole, objectively, quantitatively. Although OCT continues evolve, automated image analysis interpretation has not kept pace. With data being generated increasingly larger amount higher sampling rates, there strong need support diagnosis tracking. further amplified by fact an ophthalmologist under standard clinical conditions does have assistance specialist interpreting beforehand. A software system capable can potentially assist clinicians decisions efficiently busy daily routines. To our knowledge, been no prior work on pathology identification images, with goal directly predicting probability each given cross-sectional frame; this method be helpful especially situations where qualified readers are easily accessible. Automated images complicated three factors. First, co-existence pathologies other pathologic changes (e.g., epiretinal membrane, vitreous hemorrhage) complicate overall appearance, challenging model individually. Second, high appearance variability within hole cases, holes different widths, depths, shapes, some covered incompletely detached explicit modeling difficult). Third, measurement reflectivity tissue affected properties overlying tissues opaque media area or blood vessels around retinal surfaces will block absorb much transmitted light respectively, thus produce shadowing effects). As result these factors, attempts hand craft set features rules identify unlikely generalize well. Instead, direct encoding statistical distribution low-level training discriminative classifiers based large expert-labeled dataset may achieve more robust performance. In study, machine learning–based automatically from fovea-centered cross section SD-OCT scan was developed. Specifically, normal macula (NM) (MH), edema (ME), age-related degeneration (AMD) were identified separately through foveal center. single-frame–based serve basic component examining complete frames volume. In work, makes diagnostic suggestions solely appearances, so stand-alone interpretation. Note true diagnosis, all available information results conjunction ancillary tests) would considered together make final decision. A preliminary version presented paper.2 article significantly extends previous publication several areas: improved method, detailed labeling agreement ophthalmologists, new ground truth majority opinion consensus, evaluation original additional experiments, effect size, inconsistent labeling, performance separate testing collected after development stage, which representative future unseen data.