Ravi Kumar Samala PhD, Presenter: Nothing to Disclose

Heang-Ping Chan PhD, Abstract Co-Author: Institutional research collaboration, General Electric Company

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

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

Jun Wei PhD, Abstract Co-Author: Nothing to Disclose

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

To develop a framework utilizing the reconstructed volume and planar projection (PPJ) image of digital breast tomosynthesis (DBT) for computer aided detection (CADe) of microcalcification clusters (MCs).

With IRB approval and informed consent, DBTs of 154 subjects (307 views) were acquired with a GE prototype system at 21 projections, 30 increments, over a 600 arc. DBT with 300 arc was reconstructed using the central 11 projections to simulate narrow-angle DBT system. SART with multiscale bilateral regularization that we developed to enhance calcifications and suppress noise was used to generate DBT volume and PPJ image. 127 views with MCs were used for training and 104 views with and 76 views without MCs were used for independent testing. Multiscale calcification response (MCR) was derived from the DBT volume. Calcification candidates were extracted by iterative region growing and thresholding of PPJ image, a subset of which with high contrast-to-noise ratio (CNR) and MCR were identified as cluster centroid objects. The CNR threshold tr and decision rules for classification of true positives (TPs) and false positives (FPs) were determined adaptively based on the statistical properties of the CNR histogram for a given view. Starting from cluster centroid objects, conditional dynamic clustering forms clusters based on tr and radial distance while continuously adjusting size and centroid position. A convolution neural network (CNN) was trained to classify TPs from tissue structures and artifacts on the PPJ image. FP clusters were further reduced by the CNN response, cluster shape and combination of size, CNR and number of candidates in a cluster. The performance of the joint DBT-PPJ framework was compared to the individual CADe in DBT and PPJ using JAFROC analysis.

At view-based test sensitivities of 80 and 85%, the joint CADe resulted in 0.92 and 3.02 FPs/view. The individual DBT and PPJ CADe achieved a maximum sensitivity of 71% (3.03 FPs/view) and 79% (2.42 FPs/view). JAFROC analysis showed a significant improvement of joint CADe compared to DBT (p<0.0001) and PPJ (p=0.0022).

Joint DBT-PPJ CADe for MCs outperforms individual CADe in DBT and PPJ. 

The joint DBT-PPJ framework improves the performance of the CADe system for MCs, further improving its potential as an adjunct in radiologist’s workflow for interpretation of DBT.

Samala, R, Chan, H, Lu, Y, Hadjiiski, L, Wei, J, Helvie, M, Digital Breast Tomosynthesis: Joint Reconstruction and Planar Projection Framework for Computer Aided Detection of Clustered Microcalcifications.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14014591.html