Multi Domain Feature Fusion and Boosting Based Learning for Robust Gallbladder Ultrasound Image Classification

Gede Angga Pradipta (Department of Magister Information Systems, Institut Teknologi Dan Bisnis STIKOM Bali, Indonesia)
Pharan Chawaphan (Mechatronics Engineering Department, Faculty of Technical Education, Rajamanggala University of Technology Thanyaburi, Thailand)
Sutikno Sutikno (Department of Informatics, Faculty of Science and Mathematics, Diponegoro University, Indonesia)
Putu Desiana Desiana Ayu (Information Technology Department, Politeknik Negeri Bali, Indonesia)
Dandy Pramana Hostiadi (Department of Magister Information Systems, Institut Teknologi Dan Bisnis STIKOM Bali, Indonesia)
Made Liandana (Department of Informatics and Computer, Institut Teknologi Dan Bisnis STIKOM Bali, Indonesia)

Journal of Applied Data Sciences
Vol 7, No 3: September 2026

Accurate identification of gallbladder conditions is critically important in the medical sector, as early detection of diseases such as gallstones, cholecystitis, carcinoma, and polyps can substantially improve treatment outcomes, reduce complications, and guide timely surgical or therapeutic interventions. Existing literature on gallbladder disease classification still presents notable gaps. Most prior works rely solely on single-domain feature extraction either deep CNN-based spatial descriptors or handcrafted statistical/texture features without exploiting the complementary strengths of multi-domain feature fusion. This study addresses these gaps by proposing a hybrid framework that combines advanced preprocessing, multi-domain feature fusion, feature selection, and ensemble classification. The preprocessing pipeline applies Non-Local Means (NLM) denoising, Contrast Limited Adaptive Histogram Equalization (CLAHE), frequency-domain low-pass filtering, and Gabor filtering to enhance image quality and highlight diagnostically relevant structures. Features are extracted by fusing deep spatial descriptors from a pretrained Inception V3 network with handcrafted statistical and texture-based features, including Gray Level Dependency Matrix (GLDM) measures. Dimensionality reduction is performed using ANOVA k-best selection to retain the most discriminative attributes. The refined features are classified using LightGBM, XGBoost, Histogram Gradient Boosting, and AdaBoost, enabling a comprehensive performance comparison. Experiments conducted on the balanced UIdataGB dataset (10,692 annotated images across nine diagnostic categories) demonstrate that LightGBM, XGBoost, and Histogram Gradient Boosting achieve near-perfect performance, with accuracies exceeding 98.6% and AUC values of 0.98 across all classes, while AdaBoost shows markedly lower discriminative capability.The results suggest that gradient boosting approaches are a promising option for multi-class gallbladder disease detection, particularly when combined with multi-domain feature fusion and appropriate preprocessing and feature selection techniques.