MedConv: Convolutions Beat Transformers on Long-Tailed Bone Density Prediction

• DOI: 10.48550/arxiv.2502.00631
• Publication Date: 2025-02-01
• Authors: Xuyin Qi; Zeyu Zhang; Hongliang Zheng; Mingxi Chen; Numan Kutaiba; Ruth Lim; Cherie Chiang; Zi En Tham; Xuan Ren; Wenxin Zhang; Nevin L. Zhang; Hao Zhang; Wenbing Lv; Guangzhen Yao; Ruili Han; Kangsheng Wang; Mingyuan Li; Hongtao Mao; Li Yu; Zhibin Liao; Yang Zhao; Minh‐Son To
• Keywords: Beat (acoustics)

Bone density prediction via CT scans to estimate T-scores is crucial, providing a more precise assessment of bone health compared traditional methods like X-ray tests, which lack spatial resolution and the ability detect localized changes. However, CT-based faces two major challenges: high computational complexity transformer-based architectures, limits their deployment in portable clinical settings, imbalanced, long-tailed distribution real-world hospital data that skews predictions. To address these issues, we introduce MedConv, convolutional model for outperforms transformer models with lower demands. We also adapt Bal-CE loss post-hoc logit adjustment improve class balance. Extensive experiments on our AustinSpine dataset shows approach achieves up 21% improvement accuracy 20% ROC AUC over previous state-of-the-art methods.