Hepatic Encephalopathy Classification via Neuro-livo Features-Based Deep Tri-path Fused Neural Network

Hepatic encephalopathy (HE) is a perilous nervous problem related to liver disease, mostly in patients with cirrhosis. Early detection and intervention are vital for improving patient outcomes. Traditional methods struggle with feature selection and require manual processing, while their generalizability is limited by the absence of annotated datasets. This work proposes a novel deep tri-path fused neural network (DTFN-Net), which is a combination of three deep learning networks for HE recognition with multi-modality medical images. The tri-modal input images [computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET)] are gathered from openly accessible datasets. The liver CT images are denoised with a scalable range-based adaptive bilateral (SCRAB) filter, and the brain PET and MRI images are denoised with contrast-limited adaptive histogram equalization (CLAHE) and Gaussian adaptive bilateral (GAB) filter for removing the noisy distortions. The noise-free brain MRI and PET images are fed into the BrainNets (EfficientNet and RegNet) to extract the brain neurological features, and the liver CT images are fed into the LiverNet (MobileNet) to retrieve the liver structural features. The extracted neuro-livo (NL) features are fused, and a fully connected layer is used to classify different classes of HE cases. The experimental findings showed that DTFN-Net reaches a high accuracy of 97.87% for detecting HE in its early stages. DTFN-Net advances the overall accuracy by 3.77, 1.55, 4.87, 25.4, 2.93, and 9.35% compared to the ML-based radiomics model, OSNN, Gaze-tracker system, ML techniques, DenseNet121 + ResNet50, and multi-layer modular algorithm, respectively.