Kidney Stone Detection in Ultrasound Images Using Deep Learning

BIL443 Project – TOBB University of Economics and Technology
Author: Hakan Doğan

---

📌 Project Description

This project presents a deep learning-based approach for automatic detection of kidney stones in ultrasound images, utilizing two modern object detection architectures: Faster R-CNN and YOLOv11.

Unlike CT scans—which are costly and involve radiation—this study focuses on ultrasound (USG) as a more accessible, affordable, and non-invasive imaging method, despite its lower contrast and speckle noise challenges.

---

🧠 Models Used

🔹 Faster R-CNN

• Implemented with PyTorch and ResNet50 + FPN backbone
• Uses transfer learning and fine-tuning
• High localization precision, but slower and prone to overfitting on small datasets

🔹 YOLOv11

• Implemented via Ultralytics API and CLI
• Trained with yolo11m.pt weights
• Real-time detection capability with high inference speed and accuracy
• Better suited for noisy and low-resource environments

---

📊 Dataset

• Source: Roboflow Universe – Kidney Stone Ultrasound (ECCKD)
• 5,431 labeled ultrasound images:
  • Kidney Stone: 15,107 bounding boxes
  • Normal Kidney: 2,280 bounding boxes
• Average resolution: 640×640 pixels
• Augmentations: Horizontal/vertical flips, grayscale conversion, resizing
• Split:
  • 90% Training
  • 5% Validation
  • 5% Testing

---

📈 Results Summary

Metric	YOLOv11m	Faster R-CNN
mAP@0.5 (average)	0.789	0.684
mAP@0.5 (Kidney Stone)	0.598	0.143
mAP@0.5 (Normal Kidney)	0.979	0.450
Inference Time (ms/image)	23.3	109.87
FPS	43	9.1

---

🧪 Implementation Details

• Language: Python
• Frameworks: PyTorch, Ultralytics YOLO
• Platform: Google Colab (GPU accelerated)
• Tools:
  • Albumentations for augmentation
  • Matplotlib, Seaborn for visualizations

---

📁 Pretrained Model Weights

The trained model files can be accessed via Google Drive:
🔗 Download Trained Weights

---

🎯 Key Contributions

• First comparative YOLOv11 vs Faster R-CNN implementation for kidney stone detection in ultrasound images
• Modular training pipelines and reproducible experimentation
• Demonstrated real-time detection potential of YOLOv11
• Investigated effects of preprocessing, augmentation, and class simplification