Effect of Similarity Metrics and ROI Sizes in Featureless Computer Aided Detection of Breast Masses in Tomosynthesis

Abstract

Tomosynthesis as a technique is being developed and studied with the goal of overcoming mammography’s limitations due to overlying tissue. Various algorithms exist for tomosynthesis datasets including a novel Computer Aided Detection (CADe) algorithm using a featureless False Positive (FP) reduction stage. The goal of this project is to study the previously unexplored effects of variation of Region of Interest (ROI) sizes as well as the crucial similarity metrics for such a CADe algorithm’s performance. Four datasets consisting of 1479 tomosynthesis ROIs were generated by a CADe algorithm from reconstructed volumes of one hundred subjects consisting of 4 different sizes – 128 x 128, 256 x 256, 512 x 512, and 1024 x 1024 pixels. Five different similarity metrics – (1) mutual information, (2) average conditional entropy, (3) joint entropy, (3) Jensen divergence and (4) average Kullback-Leibler divergence were used for the task of FP reduction using a leave-one-case-out sampling scheme. Mutual information and average conditional entropy were the best performing metrics with an Area Under Curve (AUC) of 0.88. Cross-bin measures performed consistently higher than those that rely on only marginal distributions. Also, for all metrics, the datatset consisting of 256 x 256 pixel ROIs gave the best performance. In conclusion, for the tomosynthesis dataset, cross-bin measures such as MI and average conditional entropy should be used over other metrics using a ROI size of 256 x 256 pixels.