Computer-assisted diagnosis of prostate cancer using DCE-MRI data: design, implementation and preliminary results

Abstract

Objects: We present computer-assisted diagnosis (CAD) software designed to improve prostate cancer detection using perfusion MRI data.

Methods: In addition to standard visualization features, this software allows for the 2D and multislice 2D contouring of suspicious areas based on a seeded region growing algorithm, and area labeling based on zonal anatomy. Tumor volume assessment and the semiquantitative analysis of DCE-MRI sequences can both be performed. We retrospectively analyzed DCE-MRI examinations of 100 patients and found 121 lesions showing a suspiciously high intensity with early enhancement in 84 of them. Seventy-one patients turned out to be malignant, whereas 50 were benign. Based on an analysis of the median wash-in and wash-out values of these foci, we designed a standardized 5-level cancer suspicion score (ranging from "probably benign" to "highly suspicious"). This comprehensive score provides a scaled likelihood of malignancy in the region of interest taking account of its location in relation to prostate zonal anatomy. We compared its accuracy with that of visual assessments of time-intensity curves performed by specialist and non-specialist radiologists.

Results: Parameters of the scoring algorithm were designed to provide the greatest possible sensitivity in our sample population. A re-substitution evaluation provided an Se/Sp of 100/45% for peripheral zone cancer, and 100/40% for transition zone cancer characterization. When identifying malignant areas using time-intensity curves data, this simple algorithm performed significantly better (AUC = 0.77) than a non-specialist (AUC = 0.57, P < 0.0001) radiologist, and better than a trained (AUC = 0.70) radiologist, although this difference was not significant.

Conclusion: Our new prostate MRI CAD software provides a standardized cancer suspicion score for suspicious foci detected in DCE-MRI T1-w images. Our results suggest that it may improve radiologists' performances in prostate cancer identification, especially when they are not specialized in prostate imaging.