Designing Light Weight and Effective Deep Learning Model for Osteoporosis Classification

Abstract

Osteoporosis is a progressive bone disease that deteriorates bone density and structure, increasing fracture risk, particularly in older adults. While deep learning models such as ResNet-18, VGG16, MobileNetV2, EfficientNet-B0, DenseNet-121, SqueezeNet, and Shuff- leNetV2 have demonstrated promising diagnostic capa- bilities in medical imaging, their high computational com- plexity limits their applicability in real-time and resource constrained clinical settings. This paper proposes a CBAM guided structured pruning framework that intelligently compresses convolutional neural networks while preserv- ing diagnostic accuracy. By leveraging dual attention mechanisms—channel and spatial—via the Convolutional Block Attention Module (CBAM), the framework identifies and retains clinically relevant features and prunes redundant channels layer-wise. Experiments on multiclass knee X-ray datasets show that the proposed approach achieves an average parameter reduction of 50.9%, FLOP reduction of 41.3%, and inference speedup of over 40%, with minimal loss in accuracy. Notably, the pruned ResNet-18 model achieved 93.8% accuracy with a 55.9% reduction in parameters. Compared to conventional pruning techniques, the proposed method maintains clinical-grade performance across multiple architectures.