Feature-Driven EnsembleX: An Advanced Ensemble Framework for Enhanced MRI Abdomen Image Classification Using Feature Refinement and Boosting Techniques

The classification of MRI abdomen images, particularly for liver conditions, remains a complex task due to the intricate anatomy and variability of abdominal structures and the subtle imaging features that distinguish specific diagnoses. In this study, we introduce a new framework that utilizes MRI T1 water-only images generated using the Deep-Dixon technique. This technique processes Dual-sequence T1 MRI scans, comprising in-phase and out-phase images, to enhance tissue differentiation. These MRI T1 water-only images are then used for classifying liver tissue to distinguish between healthy and unhealthy liver classes. The framework begins with the extraction of detailed features, including texture, edge information, and intensity patterns from the MRI T1 water-only images. To refine the extracted features, Principal Component Analysis (PCA) was used to reduce dimensionality, followed by Recursive Feature Elimination (RFE) to select the most relevant attributes for classification. The classification is powered by an ensemble learning approach, incorporating boosting to improve weak learners and enhance model accuracy iteratively. Additionally, bagging is applied to train multiple models on bootstrapped subsets of the data, aggregating predictions to reduce variance and improve robustness. The model’s effectiveness was assessed using statistically validated performance metrics across 5-fold cross-validation, achieving an accuracy of 98.84% $\pm ~0.12$ %, a precision of 99.03% $\pm ~0.10$ %, a recall of 99.51% $\pm ~0.10$ %, specificity of 80.00% $\pm ~0.18$ %, an F1-score of 99.27% $\pm ~0.10$ %, and a ROC-AUC score of 89.76% $\pm ~0.32$ %, demonstrating its effectiveness in accurately classifying liver tissue into healthy and unhealthy categories based on MRI T1 water-only images. The evaluation was conducted on the publicly available CHAOS dataset, which demonstrates improved classification accuracy and reliability over existing state-of-the-art models, also reproducibility and validation through the use of publicly accessible data. This approach offers a promising tool for medical practitioners, facilitating precise diagnostic assessments and informed treatment planning, especially in resource-limited settings where radiological expertise may be scarce.