Quantitative Analysis Of Morphological Changes In Retinal Maculopathy

Purpose To develop and evaluate a method for automated recognition and measurement of morphology changes from optical coherence tomography(OCT)images and ulterizing multispectral imaging(MSI)of maculopathy in retinal vascular diseases.Methods OCT data from 50 patients with pathological myopia(PM),20 patients with diabetic retinopathy(DR),28 patients with retinal vein occlusion(RVO)and 50 patients with polypoidal choroidal vasculopathy(PCV)were collected and manually labeled by specialists.We propose a novel denosing method based on sparse and low-rank matrix decomposition for layer segmentation in PM and DR OCT images and compare with common used denoising methods via signal noise ratio(SNR).Then we proposed a retinal layer segmentation method based on energy minimization between layers and compare it with the build-in automatic segmentation software in commercial machine.The unsigned border position error is used to validate the segmentation accuracy.MSI images of RVO patients are ulterized to evaluate the accuracy of disease prediction.Finally,We propose a novel progressive learning framework via deep neural networks(DNN)for automated pigment epithelium detachment(PED)segmentation and compare it with state-of-the-art PED segmentation methods.Segmentation accuracy such as true positive volume fraction(TPVF),dice similarity coefficient(DSC),positive predictive value(PPV)and false positive volume fraction(FPVF)were measured.Correlation of PED volumes measured between our framework and different experts were also measured.Results Experimental results show that the proposed method performs better than other methods in the whole dataset with 22.30 SNR.The overall mean unsigned border positioning error for layer segmentation are 3.305±4.529μm in PM patients and 5.049±9.986μm in DR patients,lower than the results from build-in algorithm(14.875±44.275 μm and 14.796±45.363 μm,respectively).The ROC curve analysis show that MSI findings was a good discriminator for retinal vein occlusions than FP(AUC=0.911 vs 0.768,P=0.0318).Mean segmentation accuracy of our framework was 85.74% TPVF,85.69 % DSC,86.02% PPV and 0.38% FPVF.High correlation coefficient(r)of our segmentation results with Expert Ⅰ(r=0.9986)and Expert Ⅱ(r=0.9981)indicates that these results are highly related.Conclusions Comparing to other denosing methods,our denosing method have good performance.Our layer segmentation method is also better than the build-in algorithm,ulterizing MSI methods could improve diagnosis of RVO disease.Our proposed framework outperforms existing state-of-the-art PED segmentation methods and work with multiple types of PEDs while other methods cannot.Such superiority makes our framework a potential robotic assistant for retinal vascular diseases management and diagnosis.