Activation Function Selection for U-net Multi-structures Segmentation of End-Diastole and End-Systole Frames of Cine Cardiac MRI

Heart disease, especially Coronary Heart Disease (CHD) is a major cause of morbidity and mortality in all corners of the world. In Indonesia in particular, this disease is the highest burden of government financing. The heart as an important organ in humans has a complexity in its structure, for example, the appearance of the left ventricle will look different in two phases of heart rate, End-Diastole (ED) and End-Systole (ES). In this study, the discussion focuses on the multi-structures segmentation of the 2017 ACDC public cine cardiac MRI dataset consisting of training data from 100 patients and test data from 50 patients. Annotations made by experts are included in the training data. The Deep Learning segmentation model is built based on the 2D U-net architecture by taking into account the selection of activation functions, especially in ED and ES frames. The experimental results show that the use of ReLU activation gives better results than MISH with an accuracy of 99% for both ED and ES frames, 2% higher than ED with MISH. Furthermore, although in general the loss from multiclass segmentation is said to be good with a value range of 10 - 2, the loss due to ReLU shows better results, which is 0.02 lower than 0.04 MISH. Furthermore, the mean IoU value, quantifying the similarity between the original image and the predicted image, was found to be 0.815 for ED with ReLU and this value was 0.2 higher than ED with MISH. As for ES, the mean IoU between ReLU and MISH is not much different, which is around 0.7. As sustainability, in the next research, it is necessary to search for known method for classifying by right each part of the multi-structures heart, particularly the myocardium, in an effort to help diagnose CHD cases.