Lauren M. Kim MD, Presenter: Nothing to Disclose

Holger Reinhard Roth PhD, Abstract Co-Author: Nothing to Disclose

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

Kevin Cherry, Abstract Co-Author: Nothing to Disclose

Shijun Wang, Abstract Co-Author: Nothing to Disclose

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

Ronald M. Summers MD, PhD, Abstract Co-Author: Royalties, iCAD, Inc Research funded, iCAD, Inc Stockholder, Johnson & Johnson Grant, Viatronix, Inc

To assess the performance of a retroperitoneal lymph node (LN) computer-assisted detection (CADe) system using a novel employing Random Forest (RF) and Convolutional Neural Network (CNN) machine learning classifiers in tandem.

One radiologist, serving as the standard of ground truth, labeled 595 abdominal LN (>1cm in the short axis) on 86 CT examinations which were assorted into training (60/86), cross-validation (12/86), and test (14/86) sets. CADe was comprised of a two-phased approach, the first consisting of retroperitoneal LN candidate generation by a RF classifier which generated 40 false positives (FP) per patient at maximum sensitivity. Subsequently a CNN classifier using 100 observers was trained at 100% sensitivity with FP detections of the RF classier used as training examples of true negatives. Subsequently, the CADe was set at an operating point to display marks with confidences of at least 0.5. These CADe marks were appraised by an independent, unbiased radiologist who reviewed the 14 CT examinations in the test set and identified each identified retroperitoneal LN as undetected by CADe [false negative (FN)], CADe true positive (TP) representing a LN >8mm in the short axis, or CADe FP.

In an independent analysis by a radiologist, CADe sensitivity on the test set was 83% generating on average 2 FP per patient. No physical feature was definitively determined to elevate the FP rate except the presence of ascites (Chi-square test, p<0.05) which elevated the FP rate approximately 4-fold.

This retroperitoneal LN CADe is highly sensitive at a low FP rate. Ascites confounds this CADe system, substantially and significantly elevating its FP rate.

The accurate detection of lymph nodes plays a critical role in the diagnosis, staging, and subsequent management of neoplastic malignancy though is an inherently difficult task given their variable size, appearance, and location. Here we present a retroperitoneal LN CADe employing two machine learning algorithms in tandem which substantially outperforms previous state-of-the-art techniques which are reported to generate 3-6 FP per volume of interest (VOI) at sensitivities ranging from 53-61%. With further validation and refinement, our LN CADe may substantially bolster a radiologist’s sensitivity and proficiency in the assessment of lymph nodes.

Kim, L, Roth, H, Lu, L, Cherry, K, Wang, S, Turkbey, E, Summers, R, Performance Assessment of Retroperitoneal Lymph Node Computer-Assisted Detection Using Random Forest and Convolutional Neural Network Machine Learning Classifiers in Tandem.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14015309.html