Siam Tahsin Bhuiyan, Rubaya Khatun, Samiul Karim Mazumder, Fatin Israq, Sefatul Wasi, Rashedur Rahman, Ashraful Islam, Saadia Binte Alam
2025 IEEE International Women in Engineering (WIE) Conference on Electrical and Computer Engineering (WIECON-ECE)
In: 2025 IEEE International Women in Engineering (WIE) Conference on Electrical and Computer Engineering (WIECON-ECE)
IEEE, pp. 379-384
Pelvic fractures are often first assessed via radiograph. Yet low contrast and subtle fracture lines in radiographs contribute to missed diagnoses in up to 22% of trauma cases, while 51% of the patients remain underdiagnosed. In this study, we investigate the effects of five image preprocessing techniques, Histogram Equalization (HE), Gamma Correction, Contrast-Limited Adaptive HE (CLAHE), Image Complement, and Balance-Contrast Enhancement Technique (BCET), on the binary classification (normal vs fractured) performance of a ResNet50 convolutional neural network using the PXR150 dataset. Each preprocessing method was applied independently to create five enhanced datasets alongside the original set. Both the original and the five preprocessed datasets were trained and tested on the ResNet50 model using a 5-fold cross-validation technique. Performance metrics, including accuracy, specificity, recall, and AUROC, were computed and averaged across folds. Compared to the original baseline accuracy of 77.33% and AUROC of 0.7790, CLAHE and Gamma Correction achieved the highest accuracy of 80.67% and AUROC of 0.8140 and 0.8160. BCET also improved performance (accuracy 80.00%, AUROC of 0.8090). These findings demonstrate that localized and nonlinear contrast enhancements improve CNN-based pelvic fracture detection, underscoring the importance of tailored preprocessing in medical imaging workflows.
Assistant Professor
Department of Computer Science and Engineering
Independent University, Bangladesh (IUB)Human-Computer Interaction, AI for Social Good, AI for Public Health, AI for Impact