Evidential dempster shafer-based CNN architecture for fetal planes detection from 2D ultrasound images leveraging fuzzy-contrast enhancement and explainable AI

Abstract

In order to systematically find specific signs of development of the fetus that are present in ultrasound images, automated picture categorization in large-scale retrospective assessments can be useful for ultrasound image processing and interpretation utilizing machine learning. The advancement of automated diagnosis while preserving accuracy has greatly benefited from the use of Deep Learning architectures in medical picture analysis. The objective of this work is to precisely identify fetus planes from ultrasound images. A dataset of 12400 images is used to train the models. This study showed the effect of enhanced ultrasound image quality using FUZZY LOGIC in fetal plane detection. A proposed Evidential Dempster-Shafer Based CNN Architecture is used and the outcome is compared with PReLU-Net, SqueezeNET, and Swin Transformer. Dempster-Shafer ensures that all the pieces of evidence are properly analyzed and the classification output can be a range containing belief and plausibility rather than a single probability. Thus it handles the uncertainty. The results of each classifier were noteworthy with the proposed Evidential Classifier’s accuracy reaching 83%. The result is evaluated in terms of training and testing accuracies. Moreover, the Lime Algorithm and the Gram Cam are used to examine the classifier’s decision-making process to incorporate explainability.