Integrating YOLO-based detection into clinical workflows: A system for medical image analysis in orthanc

Details

Supervisors

Jodogne, Sébastien

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enAutomated detection of abnormalities in medical imaging is crucial for timely diagnosis and treatment planning. Deep learning models, particularly You Only Look Once (YOLO), have shown significant promise in this domain. This master’s thesis presents the development, evaluation, and integration of a modular inference system designed for medical image analysis, specifically focusing on fracture detection and pulmonary nodule identification. The system architecture incorporates three distinct image preprocessing methods—Plain Resize, a custom Tiling approach, and Slicing Aided Hyper Inference (SAHI)—to optimize image input for YOLO models of varying resolutions and object scales. Furthermore, the impact of different inference engines on performance was investigated, including PyTorch (CPU/GPU), ONNX Runtime (CPU), and NVIDIA TensorRT (GPU). For evaluation, YOLOv11 models were trained and tested. On the FracAtlas dataset for fracture detection, the model achieved a mean Average Precision (mAP@0.5) of 0.5762 and an F1-score of 0.630. For lung nodule detection on a curated subset of the LIDC-IDRI dataset, the model achieved a mAP@0.5 of 0.786 and an F1-score of 0.80. A key contribution of this work is the practical integration of the detection system as a plugin for the Orthanc Picture Archiving and Communication System (PACS) server. This plugin facilitates the application of the trained models to DICOM images, generates DICOM Structured Reports for the findings, and includes a web viewer interface, demonstrating a viable pathway for deploying AI tools within clinical imaging workflows. The thesis provides a comprehensive analysis of the trade-offs between preprocessing techniques, inference engine selection, detection accuracy, and computational efficiency, offering valuable insights for the development and deployment of AI-driven medical diagnostic aids.