DEEP LEARNING FOR LYMPH NODE DETECTION IN CT IMAGING: A BENCHMARKING STUDY OF OBJECT DETECTION FRAMEWORKS

Abstract

Accurate detection of lymph nodes on computed tomography (CT) scans is essential for cancer staging and treatment planning, yet manual identification on CT is time-consuming and subject to considerable inter-observer variability, motivating the development of reliable automated methods. Automating this task remains challenging because many nodes are very small, contrast with surrounding tissue is frequently low, and target instances constitute only a minuscule fraction of all voxels, an extreme class-imbalance problem. In this study, we evaluate five state-of-the-art object detection architectures (Faster R-CNN, YOLOv8-m, MULAN, YOLOv11-m, and RT-DETR-L) under standardized training and evaluation protocols on two datasets with different anatomical compositions: a heterogeneous multi-region dataset (Ver1) and a mediastinal-focused dataset (Ver2). Both datasets were processed using a memory-efficient 2.5D pipeline that feeds three consecutive CT slices as input, capturing volumetric context while avoiding the computational cost of full 3D models. On Ver2, Faster R-CNN and MULAN achieved the highest mAP50 of 0.641, while RT-DETR-L led in recall (0.646) and F1-score (0.630). Every model improved its mAP50 on the anatomically focused dataset. Because the same shift appeared across five structurally different detectors, we attribute it to the detection task itself rather than to any particular architectural design. Paired testing on Ver2 showed that MULAN significantly outperformed both YOLO variants in mAP50, with Faster R-CNN following the same tendency. On Ver1, no pair of models differed significantly, so performance converges once anatomical diversity increases. Taken together, the comparison gives clinicians practical grounds for choosing between speed, precision, and sensitivity, and supplies reference baselines for later work on automated lymph node detection in CT.