Demystifying evidential Dempster Shafer-based CNN architecture for fetal plane detection from 2D ultrasound images leveraging fuzzy-contrast enhancement and explainable AI

Ultrasound imaging is a valuable tool for assessing the development of the fetal during pregnancy. However, interpreting ultrasound images manually can be time-consuming and subject to variability. Automated image categorization using machine learning algorithms can streamline the interpretation process by identifying stages of fetal development present in ultrasound images. In particular, deep learning architectures have shown promise in medical image analysis, enabling accurate automated diagnosis. The objective of this research is to identify fetal planes from ultrasound images with higher precision. To achieve this, we trained several convolutional neural network (CNN) architectures on a dataset of 12400 images. Our study focuses on the impact of enhanced image quality by adopting Histogram Equalization and Fuzzy Logic-based contrast enhancement on fetal plane detection using the Evidential Dempster-Shafer Based CNN Architecture, PReLUNet, SqueezeNET, and Swin Transformer. The results of each classifier were noteworthy, with PreLUNet achieving an accuracy of 91.03%, SqueezeNET reaching 91.03% accuracy, Swin Transformer reaching an accuracy of 88.90%, and the Evidential classifier achieving an accuracy of 83.54%. We evaluated the results in terms of both training and testing accuracies. Additionally, we used LIME and GradCam to examine the decision-making process of the classifiers, providing explainability for their outputs. Our findings demonstrate the potential for automated image categorization in large-scale retrospective assessments of fetal development using ultrasound imaging.

Automated summary

Ultrasound imaging is a valuable tool for assessing fetal development during pregnancy, but manual interpretation of ultrasound images is time-consuming and subjective. Automated image categorization using machine learning algorithms can simplify the interpretation process by identifying stages of fetal development. This research aimed to improve the precision of detecting fetal planes in ultrasound images through training convolutional neural networks on a dataset of 12400 images. Results showed noteworthy performance of each classifier, with PreLUNet achieving the highest accuracy of 91.03%. LIME and GradCam were used to provide explainability for the classifiers' outputs. The findings demonstrate the potential of automated image categorization in large-scale retrospective assessments of fetal development using ultrasound imaging.