An Enhanced Residual Architecture for Automated Detection of Bone Metastasis on WholeBody SPECT Images

Bone metastases are a critical complication in oncology, demanding prompt and reliable detection to inform therapeutic strategies. This project employs the BS-80K whole-body SPECT dataset from Kaggle to develop an automated binary classifier that distinguishes normal from abnormal scans. We utilize the ResNet50 convolutional architecture, enhanced with residual connections, to improve gradient flow and representation learning for effective feature extraction. A robust preprocessing pipeline performs intensity normalization, image denoising, and data augmentation (random flips, rotations, and scaling) to enhance model generalization across heterogeneous patient data. The model is trained and validated within a Python/Flask backend, which handles image ingestion, inference, and response formatting. The user interface, implemented with HTML, CSS, and vanilla JavaScript, enables clinicians to upload SPECT images and receive classification results in real time. A comprehensive evaluation is conducted to assess predictive performance, inference latency, and computational efficiency, validating the system’s suitability for deployment in diverse clinical environments. By integrating the optimized ResNet50 architecture with a lightweight web service, this work presents a scalable solution to accelerate bone metastasis screening and support decision-making in both high-throughput centers and resource-constrained settings.

Keywords: Bone Metastasis; SPECT Imaging; ResNet50; Residual Connections; Flask; Deep Learning; Web-Based Classification.