Analysis of Machine Learning Algorithms for Real-Time Gallbladder Stone Identification from Ultrasound Images in Clinical Decision Support Systems

Global epidemiology of gallstones in the twenty-first century affects millions of individuals, and ultrasound diagnostics effectively assess gallbladder size and function and detect abnormalities. This study collected datasets from local hospitals and reliable online sources for analysis using advanced CV/IP tools and WEKA. Image preprocessing techniques, including cropping, resizing, and grayscale conversion, were applied to 90 ultrasound images, extracting 600 ROIs with 21 features spanning binary, histogram, and texture attributes. The dataset was divided into balanced training and validation subsets, and supervised learning algorithms were optimized via cross-validation and grid search. Circular patterns were processed iteratively, with specific dimensions (512 × 512 for width/height, 32 × 32 for radius/blur, 128 × 128 for columns/rows). The performance of various machine learning classifiers was evaluated using accuracy, precision, recall, F1 score, AUC-ROC, MCC, Kappa GDR, and Dice Index, ensuring strong classification of normal and abnormal samples. The random forest (RF) classifier achieved the highest performance with an accuracy of 96.33%, followed by the MLP and Logit Boost classifiers with 95.67% and 95.40% accuracy rates, respectively. The RF model also exhibited the highest precision (0.9542), recall (0.9732), F1 score (0.9636), and a Dice Index (0.9649) with an MCC of 0.925, ROC area of 0.988, Kappa (0.921), and specificity of 95.34%-indicating its strong ability to balance true positives and negatives while minimizing misclassifications. The MLP classifier also performed well with a precision of 0.9477, a recall of 0.9665, and an F1 score of 0.957, while Logit Boost had similar results with a precision of 0.9411 and a recall of 0.9665. Other classifiers, such as the Bayes Net and J48 classifiers, showed slightly lower performance with accuracy rates of 94.67% but still exhibited good precision and recall, making them viable alternatives. This study highlights that the RF classifier achieved the highest superiority among other models in detecting gallbladder stones.