Early Detection of Peritoneal Dialysis Complications Through Convolutional Neural Networks
AbstractPeritoneal dialysis is an alternative for patients with chronic renal failure requiring periodic analysis of the resulting liquid for the early detection of complications, which involves a direct evaluation of the liquid under a microscope and a biochemical test. Alternatively, the liquid could be analyzed through a photograph (indirect evaluation), enabling the early detection of complications, without requiring the patient going to a nephrology center, improving their life quality. In [Comas et al., XX Congreso Argentino de Bioingeniería, pp. 477–486 (2015)], detection of pathological samples of the liquid from photographs was preliminary studied using color descriptors and k-nearest neighbors as classifier. In the present paper, a method based on convolutional neural networks is presented, starting from Alexnet and using transfer learning. The classification phase was implemented with a multilayer perceptron, classifying the photographs between “normal” and “pathological”, using the biochemical test as Gold-standard. An error rate of 5.79%, a FPR of 4.21% and a FNR of 7.37% were obtained with great stability, reflected in low standard deviations in the estimation of the error measures. The proposed method is more robust than the previous approach, without requiring any preprocessing or feature extraction, being a good starting point for the development of an automatic tool with adequate diagnostic capacity.