Transformer-enhanced deep ensemble for multi-class liver disease classification using computed tomography images

The liver related diseases such as cirrhosis, fatty liver disease, and hepatocellular carcinoma raise significant health challenges in the world due to their increasing prevalence and their complexity in detection. This study features a deep learning-powered computer-aided diagnostic system that includes computed tomography (CT) images in the classification of liver diseases into groups of several classes. The task is to improve diagnosis quality through the implementation of all three pre-trained convolutional neural networks (CNNs) (ResNet50V2, DenseNet121, and MobileNetV2) to make multi-scale and multi-class imaging classification. All CNN models were thoroughly fine-tuned and evaluated at first. Transformer blocks were then added to every backbone to form a hybrid model. The models were trained and assessed on a CT liver dataset and performance measured based on precision, recall, F1-score and Matthews Correlation Coefficient. Findings show that there are marked enhancements with the addition of transformers especially in diagnosis of complex conditions, e.g., cirrhosis and fatty liver. The ensemble model, which was improved using the transformer, had the best overall accuracy of 97% which was higher than any single model. This study addresses the clinical benefits of CNNs, and transformers used together to classify liver diseases and provides suggestions about its application in clinical practice.