A Lightweight Hybrid Deep Learning and Adaptive Texture Encoding Framework for Accurate Hepatitis Prediction Using Ultrasound Liver Images

Hepatitis is a serious liver disease which must be diagnosed early and precisely to avoid serious complications. The application of ultrasound imaging is quite popular as it is non-invasive and inexpensive, but there are issues like the presence of speckle noise, low contrast, and intricate patterns of tissues that restrict the accuracy of the obtained diagnosis. This paper presents a lean hybrid system of automated hepatitis prediction with the use of ultrasound images. The effective pre-processing is combined with the model of Adaptive Pearson Residual Normalization of Ultrasound (APRN-U) and Contrast Limited Adaptive Histogram Equalization (CLAHE). It utilizes a dual-branch approach to feature extraction, based on deep spatial features of a lightweight Convolutional Neural Network (CNN) and Granular Texture Encoding (GTE) which incorporates GLCM, LBP, wavelet energy features, and fractal features. The features extracted are combined with an attention-based mechanism and trimmed down with the help of Principal Component Analysis (PCA). Final classification is performed using a hypertuned Bayesian Neural Network (BNN) which in turn allows uncertainty-aware prediction. As experimental data show, the proposed model has a 92% accuracy with an AUC of 0.92 and is better than current approaches. The contribution of each component is further authenticated in the ablation study. The suggested framework provides a computationally effective and valid solution to detect hepatitis in clinical cases.