Artificial Intelligence Diagnosis System For Age-related Macular Degeneration And Polypoidal Choroidal Vasculopathy

BACKGROUNDAge-related macular degeneration(AMD)is one of the most important age-related retinal diseases for vision loss.The main pathological changes are characterized by dengeration of the macular chordoial capillaries,Bruch membrane and retinal pigmental epithelium(RPE),which further cause central vision impairment.AMD remains to be the third main disease for irreversible blindness around the world,population studies show that there will be over 200 million patients diagnosed with AMD in 2020.Polypoidal choroidal vasculopathy is the polypoidal dilation at the end of branch vascular network(BVN)of the choroid,which is now recognized as one subtype of AMD.Recent clinical studies demonstrate that PCV comprised of 20-60%among Asian patients with neovascular AMD,whilst the percentage decreased to 8-13%in Caucasion patients.Differences can be seen in clinical strateties,diseases progress and prognosis between AMD and PCV.OBJECTIVE:To establish and train deep learning convolutional neural network models using fundus color photographs and spectral-domain optical coherence tomography(SD-OCT)images,and to realize classification of AMD(dry-AMD,wet-AMD,PCV,normal).The secondary goal is to compare the diagnosis efficiency between bi-modal and uni-modal artificial intelligence models.Performance levels between retinal specialists and the best deep learning algorithms are further compared.METHODS:We retrospectively collected cases from the ophthalmology outpatient clinic of Peking Union Medical College Hospital(PUMCH).The inclusion criteria included patients who were diagnosed with AMD or PCV,and had images of the color fundus,SD-OCT,and angiography.Patients with concurrent diseases as glaucoma or diabetic retinopathy were excluded.The normal controls were collected from the Physical Examination Medical Center of PUMCH with normal color fundus images and SD-OCT images.All fundus images from patients and the normal controls included were collected under the same criteria.The clearest image of the posterior pole of the color fundus was selected and saved with the jpeg.format and the standard denomination.The same collection strategy was applied to the relatively matched SD-OCT images.Preprocessing was done towards all images to conceal the diagnosis tag and to standardize the image format before training.The picture dataset was further sorted into 3 different subsets for training,validation and test automatically,every single image only existed in one of the subsets.Appropriate deep learning models from the repertoire were selected and moderated for our novel best-fit machine learning systems.The same test subset was used to evaluate all models built as well as retinal specialists,to further compare the performance levels of artificial intelligence systems and retinal specialists.RESULTS:Four bi-modal and 8 uni-modal machine learning models were built and evaluated in this study.Color fundus images from 1099 eyes and SD-OCT images from 821 eyes were collected respectively.Among all machine learning models and human experts,our best bi-modal deep learning convolutional neural network(CNN)model won the best efficacy of classification,with the accuracy 87.4%,sensitivity 88.8%,specificity 95.6%.And the agreement with the diagnosis gold standard was higher than that of retinal specialists.Furthermore,its ability to categorize PCV from all fundus images tested was far better than any other models as well as retinal specialists.CONCLUSION:This study designed a novel bi-modality deep learning CNN model using color fundus and SD-OCT images to classify PCV and AMD.Its high accuracy and agreement with the diagnosis gold standard reveals a promising application future in the realm of public health.