Long Short-Term Memory Model Based Microaneurysm Sequence Classification in Fundus Images

Diabetic Retinopathy (DR) has emerged as one of the serious medical conditions over the years leading to blindness among patients. Microaneurysms (MAs) are generally the earliest objective evidence of DR captured in fundus imaging. This work proposes a novel methodology based on long short-term memory (LSTM) to exploit the sequence dependencies of 1-D feature signals extracted from MAs and aid in their classification in colour fundus images. The model is trained using 1-dimensional intensity based signals generated from various patches of preprocessed fundus images. The model is tested on e-ophtha & ROC datasets and sensitivity scores are computed against seven unique values of false positive per image. The average of these scores is utilized as performance measurement of the proposed model which shows 66.6% and 60.5% sensitivity for e-ophtha and ROC datasets, respectively.