Lightweight and Accurate YOLOv7-Based Ensembles With Knowledge Distillation for Urinary Sediment Detection
Keita Sasaki, Hiroki Nishikawa, Ittetsu Taniguchi, Takao Onoye · IEEE Access · 2025
Urine sediment analysis plays an important role in evaluating kidney function. In addition to improving detection accuracy, reducing model size is also a key challenge, especially when considering deployment on medical devices where computational resources are limited. To address these demands, we propose a lightweight and accurate detection framework that combines YOLOv7-based ensemble learning with feature-based knowledge distillation. In our approach, two attention-enhanced YOLOv7 variants are used as teacher models, which transfer their feature representations to compact YOLOv7-tiny student models. These student models are then integrated using weighted boxes fusion to further enhance detection performance. Evaluated on a real-world urinary sediment dataset, the proposed framework achieves 84.6% precision, 88.1% recall, and a [email protected] of 0.920. These results represent an improvement of more than 3% over the baseline YOLOv7-tiny and slightly outperform the state-of-the-art YOLOv7-CBAMteacher model. Moreover, the model size is reduced to approximately one-third ofYOLOv7-CBAM, significantly improving its suitability for deployment in resource-constrained environments. Ablation studies further confirm the complementary strengths of the ensemble design and suggest that additional performance gains may be achievable through more adaptive fusion strategies.