Jun Wei PhD, Presenter: Nothing to Disclose

Heang-Ping Chan PhD, Abstract Co-Author: Nothing to Disclose

Lubomir M. Hadjiiski PhD, Abstract Co-Author: Nothing to Disclose

Chuan Zhou PhD, Abstract Co-Author: Nothing to Disclose

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

Mark Alan Helvie MD, Abstract Co-Author: Institutional grant, General Electric Company Consultant, General Electric Company

To design a multi-channel (MCH) method for characterization of true microcalcifications (MCs) and false positives (FPs) in a computer-aided detection (CAD) system for digital breast tomosynthesis (DBT).

With IRB approval and informed consent, CC- and MLO-view DBT of 40 breasts containing biopsy-proven MCs were acquired with a prototype GE GEN2 DBT system (21 projection views (PVs), 60º arc, 3º increments). The CAD system first prescreened for MC candidates in each PV using the difference-image technique and iterative thresholding. The MC candidates were refined with a local SNR-based segmentation method. An MCH approach was designed to extract the signal response from each candidate. Using the geometrical information of the DBT system, the MC candidates on all PVs were ray-traced to the breast volume. The 3D location of each MC candidate was determined as where the maximum number of rays intersect and the MC candidates on the PVs corresponding to the same signal in the breast volume were identified by the intersecting rays. The MCH responses from the corresponding MC candidates were combined and used as a decision variable to classify true MCs from FPs. We compared three types of channelized basis functions: Laguerre-Gauss (LG), Gabor, and difference of Gaussians (DoG), for the MCH method in this study. To evaluate the MCH method, the data set was randomly split by case into two independent subsets for two-fold cross-validation training and testing. The test performance was assessed by receiver operating characteristic (ROC) analysis.

An average of 147.1 MC candidates per PV was detected at prescreening. After ray tracing, a total of 13219 signals including 496 true MCs and 12723 FPs were identified in 3D (159.0 FPs/DBT volume). The area under the ROC curve (Az) for the two test subsets were 0.877±0.011 and 0.878±0.012 for LG, 0.858±0.012 and 0.862±0.013 for Gabor, and 0.874±0.011 and 0.869±0.013 for DoG basis functions. The three types of basis functions provided comparable MCH response (p>0.05).

The MCH method is promising for classification of MCs and FPs. Further study is underway to optimize the basis functions using larger training and test sets.

CAD has been found to be useful for MC detection in mammography. With DBT emerging as a screening tool, it is important to develop an effective CAD system for MCs in DBT.

Wei, J, Chan, H, Hadjiiski, L, Zhou, C, Lu, Y, Helvie, M, Microcalcification Detection in Digital Breast Tomosynthesis (DBT): Multi-channel Response for Characterization of True and False Positives.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11034414.html