Deep Prototype-Calibrated Learning for Multi-Class Assessment of Osteoporotic Status Using Knee Radiographs
6th International Conference on Emerging Smart Technologies and Applications, eSmarTA 2026, Dhamar, Yemen, 4 - 05 Ağustos 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/esmarta70636.2026.11652142
- Basıldığı Şehir: Dhamar
- Basıldığı Ülke: Yemen
- Anahtar Kelimeler: Deep Learning, Disease Detection, Knee X-ray, Osteoporosis Diagnosis, Prototype-Driven Learning
- Atatürk Üniversitesi Adresli: Evet
Özet
Knee Osteoporosis is a global health issue affecting millions of human beings, particularly elderly people. Early and accurate diagnosis of knee Osteoporosis is vital for maintaining quality of life and minimizing health complications. Recent advancements in computer vision have yielded promising results in medical image analysis; however, conventional methods often use standard classification heads, which may perform well on clear cases but struggle with visually overlapping categories, especially on small datasets. This paper presents a dual-path deep learning method enriched with a prototype-guided calibration strategy, reinforcing the discriminative consistency of the learned feature space and improving the robustness of the final decision. It integrates a main classification branch and a prototype-based calibration branch, whose outputs are fused at the logit level to classify knee bone mineral density status on knee X-ray into three categories (normal, osteopenia, and osteoporosis). The initial results on a multi-class, multi-source dataset are promising. During holdout evaluation, the proposed method obtained higher diagnostic accuracy than competing pre-trained deep learning architectures, including the Swing transformer and aggregated residual networks, achieving an F1 score (92.36%) with +2% gain over the nearest competitive model. Under 10-fold cross-validation, the proposed method maintained competitive performance relative to the strongest baseline, with comparable or slightly improved performance. Furthermore, the ablation study showed that removing prototype-based fusion clearly degraded overall performance, confirming the real contribution of this component to the model's behavior.