A Robust Multi-Stage Framework for Noise-Resilient Bone Tumor Detection, Segmentation, and Classification

Abstract: Accurate identification and segmentation of bone tumors from medical images are essential for early diagnosis, treatment planning, and prognosis estimation. However, automated analysis remains difficult because bone lesions show considerable variation in shape, size, texture, anatomical location, and boundary definition, while medical images are frequently affected by noise and low contrast. This research work presents a three-phase framework for bone tumor analysis. In Phase 1, image normalization, noise reduction, contrast enhancement, adaptive superpixels, and stochastic clustering based on Gaussian Mixture Models and Expectation-Maximization are used for noise-resilient candidate- region segmentation. In Phase 2, a Fast Mask R-CNN architecture with a VGG-19 backbone is employed for simultaneous tumor localization, segmentation, and benign/malignant classification. In Phase 3, Snake Swarm Optimization (SSO) is integrated to refine segmentation boundaries and optimize key model parameters. This results show progressive improvement across phases: the Phase 1 model achieved 96.3% accuracy, 90.5% sensitivity, 97.89% Dice coefficient, and 96.86% Jaccard index; the Phase 2 model achieved 96.31% accuracy, 94.25% precision, 97.82% recall, and 95.99% F1-score; and the optimized Phase 3 model achieved 97.8% accuracy, 97.2% precision, 97.9% recall, and 97.5% F1- score. Collectively, the findings suggest that combining adaptive segmentation, instance segmentation, and swarm-based optimization can improve robustness and clinical usefulness in computer-aided bone tumor diagnosis.

Keywords: bone tumor; medical image segmentation; Fast Mask R-CNN; adaptive superpixels; Gaussian mixture model; expectation-maximization; Snake Swarm Optimization; computer-aided diagnosis.