Tri-plane correlation imaging for the detection of breast cancer: Effects of angular separation and correlation rule

Abstract

Objective: Overlapping anatomy in projection mammography can degrade the detection of mammographic lesions and further create mass-like features that can confound computer-aided detection (CADe) systems and degrade their performance. Tri-plane correlation imaging (TCI) aims to reduce this impact by incorporating three projection images into CADe using a geometric correlation scheme. This study aimed to assess the efficacy of TCI for breast cancer detection.

Materials and Methods: An image set of 198 human subjects was used from a breast tomosynthesis database. Each case included 25 projections acquired within a 45 degree angular span at an approximate total glandular dose equal to that of two-view mammography. Triplet images were derived from each case (the central projection along with two symmetrical images) and analyzed by two independent CADe programs. The CADe results from the three projections from each program were combined using two TCI correlation rules based on unanimous vote and majority vote. The findings were analyzed in terms of true positives (TP), sensitivity, false positives per breast volume (FP), and a positive predictive index combining the two figures (i.e., TP/(TP+FP)).

Results: For the first CADe program, the TCI scheme using the majority voting rule improved sensitivity by 40% while maintaining specificity, leading to a 40% improvement in the PPI performance. For the second, a higher-sensitivity/lower-specificity CADe program, the TCI using the majority voting rule improved sensitivity by 10% while increasing false positives per breast volume by 3%, leading to an improvement of 8% in the PPI performance. The unanimous voting rule led to notably lower performance for both programs. For both CADe programs, an angular separation of six degrees (± 3 degrees) proved optimal.

Conclusions: TCI was able to improve sensitivity over single projection imaging while maintaining specificity for both CADe programs, which suggests its potential as a supplement to standard mammography and as a complementary module to existing CADe algorithms.