Error investigation of models for improved detection of masses in screening mammography

This study analyzes the performance of a computer aided detection (CAD) scheme for mass detection in mammography. We investigate the trained parameters of the detection scheme before any further testing. We use an extended version of a previously reported mass detection scheme. We analyze the detection parameters by using linear canonical discriminants (LCD) and compare results with logistic regression and multi layer perceptron Neural Network models. Preliminary results suggest that regression and multi layer perceptron Neural Network showed the best Receiver Operator Characteristics (ROC). The LCD analysis predictive function showed that the trained CAD scheme performance can maintain 99.08% sensitivity (108/109) with false positive rate (FPI) of 8 per image with ROC Az= 0.74±0.01. The regression and the multi layer perceptron Neural Network ROC analysis showed stronger backbone for the CAD algorithm viewing that the extended CAD scheme can operate at 96% sensitivity with 5.6 FPI per image. These preliminary results suggest that further logic to reduce FPI is needed for the CAD algorithm to be more predictive.