An Ensemble Deep Learning Approach for Accurate Urinary Sediment Detection Using YOLOv9e and KD-YOLOX-ViT
<p dir="ltr">Urinary Tract Infections (UTIs) are common medical diseases that occur from bacterial infections which can impact any area of the urinary system. Detecting urine particles is crucial for diagnosing UTIs and other renal abnormalities, as the presence of red blood cells (R...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| مؤلفون آخرون: | , , , |
| منشور في: |
2025
|
| الموضوعات: | |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | <p dir="ltr">Urinary Tract Infections (UTIs) are common medical diseases that occur from bacterial infections which can impact any area of the urinary system. Detecting urine particles is crucial for diagnosing UTIs and other renal abnormalities, as the presence of red blood cells (RBCs), white blood cells (WBCs), casts, fungi, and bacteria in urine can indicate disorders such as hematuria, kidney stones, and urinary tract malignancies. This study introduces an advanced deep learning method that utilizes ensemble of YOLOv9e and KD-YOLOX-ViT models. YOLOv9e model is built upon the Generalized Efficient Layer Aggregation Network (GELAN) and Programmable Gradient Information (PGI). The KD-YOLOX-ViT model integrates knowledge distillation and Vision Transformer (ViT) modules within the YOLOX framework to improve performance. This ensemble framework combines the strengths of YOLOv9e and KD-YOLOX-ViT models to automate the detection and classification of seven distinct urine sediment particles, effectively addressing challenges related to class imbalance, image resolution, and domain adaptation. The YOLOv9e model exhibited exceptional performance, obtaining precision, recall, and mean average precision (mAP50) scores of 88.5%, 88.1%, and 92.2%, respectively. The KD-YOLOX-ViT model also performed well, achieving precision, recall, mAP50, and mAP50-95 scores of 86%, 88%, 86.7%, and 53.3%, respectively. By ensembling YOLOv9e and KD-YOLOX-ViT using Weighted Box Fusion (WBF), the model achieved a final mAP50 of 94.18%. The model’s adaptability was further validated through rigorous external validation on a novel dataset, yielding in a mAP50 of 94.64%. Comparative analysis with state-of-the-art methods confirms the model’s real-time analytical capabilities. Moreover, the integration of eXplainable AI (XAI) enhances interpretability, offering valuable insights for confident clinical diagnosis. This comprehensive approach shows significant promise in advancing diagnostic accuracy and enabling earlier, real-time treatment on a global scale</p><h2>Other Information</h2><p dir="ltr">Published in: IEEE Access<br>License: <a href="https://creativecommons.org/licenses/by/4.0/deed.en" target="_blank">https://creativecommons.org/licenses/by/4.0/</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1109/access.2025.3577959" target="_blank">https://dx.doi.org/10.1109/access.2025.3577959</a></p> |
|---|