Ordinal-Aware Knee X-ray Osteoporosis Status Classification with Hard-Sample Reweighting and Multi View Inference
7th International Congress on Engineering and Life Science (ICELIS), Tbilisi, Gürcistan, 22 - 24 Eylül 2026, ss.142-151, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.61326/icelis2026tbilisi
- Basıldığı Şehir: Tbilisi
- Basıldığı Ülke: Gürcistan
- Sayfa Sayıları: ss.142-151
- Atatürk Üniversitesi Adresli: Evet
Özet
Osteoporosis is a persistent skeletal condition that weakens bone density and increases the risk of fragility fractures. This disease can progress silently for a long period and may not be detected until a low-trauma fracture develops. Fractures can cause permanent pain, functional disabilities, limited mobility, loss of independence, decreased quality of life, and a significant healthcare burden. Thus, recognizing patients at greater fracture risk, directing prompt preventive or therapeutic care, and decreasing the long-term clinical implications of osteoporosis. In this context, automated osteoporosis detection from knee radiographic images may provide supportive screening information and assist clinicians in improving early diagnostic decision-making. Although current deep learning techniques have delivered promising results, many traditional screening methods do not show high accuracy due to the easy-to-classify sample dominance problem. These models can effectively learn discriminative patterns from easily classified samples, but they may struggle with difficult samples that exhibit subtle or ambiguous features. To mitigate this problem, we present a deep learning-based method that classifies bone mineral density status from knee radiographs into three categories (Normal, Osteopenia, Osteoporosis). This method employs an ImageNet pre-trained ConvNeXt-Tiny backbone network and an enhanced training procedure using Exponential Moving Average (EMA) of the model's weights, Adaptive Hard Sampling (AHS), ordinal-aware soft labeling, and multi-view inference. By maintaining the exponential moving average of model weights throughout training, we obtained a more stable EMA model, which reduced performance fluctuations on the evaluation set and made the results more consistent. Based on EMA confidence, Adaptive Hard sampling prevents the model from getting stuck on easy examples by reweighing difficult examples and selecting them more frequently, thereby improving generalization. Ordinal-aware soft labeling to reflect the ordered nature of classes and models the uncertainty between neighboring ranks, and reduces errors between close classes. Finally, a multi-view inference approach was used to reduce errors originating from a single view by averaging predictions obtained from different crops of the same image in the test set. The initial results on a public knee X-ray dataset comprising 1,800 images 143 show that the model achieves higher diagnostic accuracy than baseline models, achieving 92.05% F1 score.