Enhanced Fetal Brain Abnormality Classification Via Position Encoded Layer Based Deep Convolutional Neural Network

Abstract

In order to diagnose and treat fetal brain disorders early, sophisticated and precise diagnostic technologies are required. The detection of minor fetal brain abnormalities (FBA) is still a challenging undertaking, even with medical imaging advancements. False positives or negatives could result from current model inability to handle patterns. Improved accuracy and reliability in fetal brain abnormalities classification is the objective of this work, which aims to address this issue. Initially, the input MRI (Magnetic Resonance Imaging) fetal brain images are collected from Kaggle. In pre-processing, the input images are segmented using Edge Stop Functional Iterative Region Growing (ESFIRG) algorithm by iteratively grouping pixels based on predefined criteria while considering edge information and effectively separate regions in an image. After, Augmentation is applied to increase the images from 172 to 3440 for further classification. Furthermore, the attributes like Mean Pixel Intensity, Median Pixel Intensity, Pixel Intensity Deviation, Label, and Class Name are extracted. From the extracted attributes the important attributes are selected by using Color Gray Coherent - Minimum Redundancy Maximum Relevance (CGC-MRMR) algorithm for reduce the complexity and increase accuracy of the classification process. Concerning the incorporation of PEL into the detection of fetal brain normal and abnormalities, there is a knowledge vacuum in the current literature. In most cases, researchers either stick to tried-and-true techniques or fail to fully utilize deep learning. This research fills that need by presenting a new PEL-DCNN method, which overcomes all the problems with the existing approaches. By combining the strengths of PEL for feature extraction and DCNN for accurate classification, the work intends to overcome the current shortcomings in fetal brain abnormality identification. Finally, the FBAs are classified by the help of Position Encoded Layer based Deep Convolutional Neural Network (PEL-DCNN) which is classify the presence of the normal or abnormal of predicting in the FBA. A DCNN is used to classify the retrieved features robustly. By combining PEL and DCNN, this method improves the model capacity to detect intricate patterns in brain images. With an F1-score of 93.75%, it accomplishes a recall of 92.31%, specificity of 97.09%, sensitivity of 92.31%, precision of 95.24%, and accuracy of 96.32%. These findings demonstrate that the effectiveness of the proposed method in tackling the difficulties of fetal brain abnormalities diagnosis.