Developed a machine learning pipeline to localize medical devices in chest x-rays with 97% accuracy across diverse test sets. This project combines a Multi-Label Classifier with a YOLOv5 model to detect and generate bounding boxes around medical devices, ensuring robust performance in varied clinical imaging scenarios.
Project Overview
The objective of this project was to enhance the detection and localization of medical devices within chest x-rays—a critical task for clinical diagnostics and effective patient care. By leveraging advanced machine learning techniques, the system not only classifies but also pinpoints the exact location of devices, streamlining the radiologist's workflow.
Key Features
- High Accuracy: Achieved 97% detection accuracy across multiple diverse datasets.
- Dual-Model Architecture: Innovative combination of Multi-Label Classifier and YOLOv5 for precise localization.
- Advanced Visualization: Automated generation of bounding boxes around detected medical devices.
- Robust Performance: Adaptable system that maintains accuracy across various imaging conditions.
Technical Details
The pipeline starts with a Multi-Label Classifier that pre-screens x-ray images to identify potential regions of interest. A YOLOv5 model is then employed to accurately localize and draw bounding boxes around detected devices. This dual-model approach ensures robustness and high precision, even in the presence of imaging variability.
Internship Insights: Radiology and ResNets
During my summer internship with AIMI at Stanford, I gained valuable insights into the application of deep convolutional neural networks—specifically ResNets—in radiological image analysis. By fine-tuning a ResNet model for chest x-ray analysis, our team was able to further improve detection accuracy and better understand the nuances of clinical imaging. This experience validated the potential of integrating state-of-the-art deep learning techniques into medical imaging workflows and reinforced the importance of adaptive models for handling diverse datasets.
You can read more about these insights on this Medium article.