Ari Seff, Abstract Co-Author: Nothing to Disclose

Evrim Bengi Turkbey MD, Abstract Co-Author: Nothing to Disclose

Le Lu PhD, Abstract Co-Author: Nothing to Disclose

Ronald M. Summers MD, PhD, Presenter: Royalties, iCAD, Inc Research funded, iCAD, Inc Stockholder, Johnson & Johnson Grant, Viatronix, Inc

To develop a new algorithm paradigm for detection of mediastinal and abdominal enlarged lymph nodes (>10 mm in short axis diameter).

Two CT datasets (90 patients with 388 mediastinal LNs and 86 patients with 595 abdominal LNs) were used in this study. Images were acquired in the portal venous phase with a slice thickness of 1-1.25 mm. All enlarged lymph nodes in the two target regions were marked with a centroid by a radiologist as the reference standard. We applied a new algorithm paradigm of aggregating lymph node detections on 2D views. First, candidate generation within the target region proceeded via a random forest classifier trained with primitive, voxel-level features. Each candidate was cropped as a cube VOI of 45×45×45 voxels. 2D slice sampling along the axial, coronal, and sagittal axes of a VOI resulted in 27 image views per candidate. Histogram of Oriented Gradients (HOG) features were extracted from each 2D view and used to train a robust LibLinear classifier. Following testing, we aggregated the 27 resulting confidence scores per candidate via max-pooling and sparse linear fusion schemes to obtain a final probability score per VOI. Six-fold cross-validation was used in the experiments.

The sensitivities of our automated detection systems were 78.0% at 6 false positives/volume (FP/vol.) (86.1% at 10 FP/vol.) and 73.1% at 6 FP/vol. (87.2% at 10 FP/vol.) for the mediastinal and abdominal datasets respectively. Approximately 20% of the false positives were actually small lymph nodes (<10 mm).

We validated a novel approach to automated lymph node detection in CT images that significantly outperforms the previous best reported work.

Detection of lymphadenopathy is crucial in cancer patients to assess staging and treatment response. Automated detection may permit more accurate and time efficient assessment.

Seff, A, Turkbey, E, Lu, L, Summers, R, Automated Lymphadenopathy Detection by Sparse Linear Fusion of 2D Views.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14011748.html