Universal lymph node detection in T2 MRI using neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate identification of lymph nodes (LNs) in T2 Magnetic Resonance Imaging (MRI) is crucial for staging cancer and assessing lymphadenopathy.
• Previous lymph node detection methods were limited to specific anatomical regions, necessitating a universal approach for broader clinical application.

Purpose of the Study:

• To develop and evaluate a Computer Aided Detection (CAD) pipeline capable of universally identifying lymph nodes in T2 MRI studies.
• To enable the detection of both benign and metastatic lymph nodes across various body regions.

Main Methods:

• Development of a CAD pipeline utilizing various neural network architectures, including Faster RCNN, FCOS, FoveaBox, VFNet, and Detection Transformer (DETR).
• Training and comparison of different models, with a focus on VFNet with Adaptive Training Sample Selection (ATSS) and ensemble methods.
• Evaluation of model performance using mean Average Precision (mAP) and recall metrics on a test set of 122 studies.

Main Results:

• The VFNet model achieved a 51.1% mAP and 78.7% recall at 4 false positives (FP) per volume.
• A one-stage model ensemble reached a 52.3% mAP and 78.7% sensitivity at 4 FP per volume.
• VFNet and the ensemble model demonstrated comparable performance, suitable for interchangeable use in the CAD pipeline.

Conclusions:

• The developed CAD pipeline successfully achieved universal detection of lymph nodes in T2 MRI.
• The system demonstrated a significant sensitivity improvement of approximately 14% over existing methods (78.7% at 4 FP vs. 64.6% at 5 FP per volume).
• This universal detection capability holds promise for enhanced cancer staging and lymphadenopathy assessment.