Computer-aided diagnosis of metabolic dysfunction-associated steatotic liver disease using paired liver–kidney texture features

This study aims to introduce a computer-aided diagnosis tool that leverages texture analysis and machine learning techniques to support radiologists in diagnosing fatty liver through ultrasound imaging. In this method, regions of interest for further analysis are identified from the paired liver-kidney. Textural features are extracted from these regions using conventional methods (e.g., GLCM) and advanced techniques (e.g., ELBP, NASTA). A two-stage feature reduction process is applied, beginning with a statistical t-test, followed by dimensionality reduction methods including PCA, SBS, SFS, LDA, and ReliefF. Finally, three classifiers: RF, SVM, and KNN are employed in two sequential stages to classify healthy versus fatty livers and grade fat accumulation severity. Furthermore, we developed a graphical user interface. We analyzed 189 ultrasound images from 54 patients. The ReliefF–SVM model in the first stage and the LDA–SVM model in the second achieved the highest performance, with an overall accuracy of 0.945 ± 0.023, precision of 0.935 ± 0.033, sensitivity of 0.940 ± 0.034, and specificity of 0.931 ± 0.036. Results demonstrate that extracting textural features from the paired liver-kidney regions using advanced methods with conventional features, and applying a two-stage binary classification framework, can significantly enhance the diagnostic efficacy of fatty liver.