Enlarged lymph nodes (LNs) can provide important information for cancer\ndiagnosis, staging, and measuring treatment reactions, making automated\ndetection a highly sought goal. In this paper, we propose a new algorithm\nrepresentation of decomposing the LN detection problem into a set of 2D object\ndetection subtasks on sampled CT slices, largely alleviating the curse of\ndimensionality issue. Our 2D detection can be effectively formulated as linear\nclassification on a single image feature type of Histogram of Oriented\nGradients (HOG), covering a moderate field-of-view of 45 by 45 voxels. We\nexploit both simple pooling and sparse linear fusion schemes to aggregate these\n2D detection scores for the final 3D LN detection. In this manner, detection is\nmore tractable and does not need to perform perfectly at instance level (as\nweak hypotheses) since our aggregation process will robustly harness collective\ninformation for LN detection. Two datasets (90 patients with 389 mediastinal\nLNs and 86 patients with 595 abdominal LNs) are used for validation.\nCross-validation demonstrates 78.0% sensitivity at 6 false positives/volume\n(FP/vol.) (86.1% at 10 FP/vol.) and 73.1% sensitivity at 6 FP/vol. (87.2% at 10\nFP/vol.), for the mediastinal and abdominal datasets respectively. Our results\ncompare favorably to previous state-of-the-art methods.\n