Deep learning model for automatic differentiation of EMAP from AMD in macular atrophy – Scientific Reports

GA is responsible for 20% of cases of legal blindness in AMD patients. GA areas are visible on fundus examination as sharply demarcated hypopigmented areas, in which larger choroidal vessels may become visible owing to the absence of the RPE and the choriocapillaris. In AMD related GA, peripheral extension of the atrophy is typically faster and patients maintain good central vision until foveal involvement, which occurs at a later stage. Histopathologic investigation of these areas showed cell death in the RPE, outer neurosensory retina, and choriocapillaris and modern in vivo technology allowed confirmation of these pathological elements. SD-OCT shows choroidal signal enhancement (due to loss of absorbing pigment) and thinning and disruption of the outer retinal layers, including the outer nuclear layer. Atrophy of the RPE and loss of fluorophores induce complete loss of autofluorescence signal in these areas at FAF imaging, which is accompanied by a variable pattern of hyper autofluorescence in adjacent areas⁸. In 2009, Hamel et al.¹ described a cohort of 18 patients in their 5th decade (age range 41–54) showing a specific GA phenotype (which is now known by the name of EMAP) characterized by peripheral degeneration symptomatic for difficulties in dark adaptation and later onset of central scotoma in the context of macular atrophic degeneration. Subtle differences may be noticed in atrophy shape: in AMD, atrophy is circular, centered on the fovea with a larger horizontal axis whereas in EMAP atrophy is oval, polycyclic with a larger vertical axis¹. Differently from AMD patients these patients do not develop choroidal neovascularization and are instead characterized by peripheral retina disfunctions. Since peripheral symptoms such as dark adaptation impairment and visual field reduction often lie unnoticed to patients’ awareness until late stages, the first symptom leading to medical referral in EMAP patients is often the appearance of central scotoma. Given the presence of GA at the posterior pole, these patients are likely to be misdiagnosed with the much more prevalent GA related to AMD, since subtle differences in sex, age and inflammatory profile are often not sufficient to raise suspect of an EMAP case. In fact, small series have shown that EMAP seem to most frequently affect women and is associated to eosinophilia, lymphocytosis, increased erythrocyte sedimentation rate, decreased CH50, and high plasma C3 level which may suggest an association between EMAP and a systemic inflammatory profile⁹. By contrast, other authors didn’t detect a blood inflammatory profile in EMAP patients¹⁰. Interestingly, in our large cohort, no significant difference in sex prevalence and age was noted between AMD and EMAP patients. An additional element of confusion is given by the fact that EMAP cases are characterized by the presence of pseudodrusen and often show a “diffuse-trickling” FAF pattern, which is typically related to faster progressing GA phenotypes in AMD¹⁰. Altogether, these elements account for a high risk of mismanagement of EMAP cases diagnosed as AMD cases. Despite the presence of a number of studies^3,4,11,12 detecting and classifying GA with deep learning methods, our study is the first in literature to show results of an automatic differential diagnosis between GA caused by AMD and EMAP cases. Our findings reveal good performance of the FAF imaging based deep learning classifier presented in the study, showing a sensitivity of 84.6% and a specificity of 85.3% for the diagnosis of EMAP on 30° × 30° FAF images and a sensitivity of 90% and a specificity of 84.6% for the diagnosis of EMAP on 55° × 55° FAF images. Interestingly, wide field images allow a significantly higher sensitivity of the DL classifier, suggesting peripheral alterations in the FAF pattern that still lay unnoticed to human eye detection. A FAF imaging-based DL classifier has recently been evaluated for the distinction between GA from AMD and atrophy due to STGD1 and Pseudo-Stargardt multifocal pattern dystrophy. This model achieved a training accuracy of 0.89 on the test set with the model training with 100 epochs and 0.92 using 10 epochs and an AUC-ROC of 0.981⁴. Besides FAF, OCTA could help us to differentiate GA from EMAP with specific defect. Indeed Rajabian et al. found three retinal regions in EMAP disease corresponding to progressively deeper perfusion defects. Quantitative analyses of retinal vessels revealed significant alterations, especially in the DCP and CC, in both atrophic and junctional zones in retina of EMAP patients compared with preserved zones and controls¹³.

Among the strengths of the study, it should be mentioned the validation of the DL classifier on an external set of images and thus the multicentricity of the investigation. Limitations of the study include the lack of comparison with DL methods based on different enface images or B scan methods and the lack of inflammatory profile testing in the study population.

In conclusion, our model could provide an automatic easy to use tool for differentiation of EMAP from early onset AMD in young patients showing macular atrophy. This could allow correct clinical management from the earliest phases, representing a valid tool in the setting of personalized medicine.