Ethan Leng1, Joseph Koopmeiners2, Lin Zhang2, and Gregory John Metzger1
1Center for Magnetic Resonance Research, Minneapolis, MN, United States, 2School of Public Health, Division of Biostatistics, University of Minnesota, Minneapolis, MN, United States
1Center for Magnetic Resonance Research, Minneapolis, MN, United States, 2School of Public Health, Division of Biostatistics, University of Minnesota, Minneapolis, MN, United States
We developed a two-stage classification model for simultaneous detection of prostate cancer on prostate MRI and localization of aggressive, high-grade PCa, using both quantitative MRI and radiomic features.
Figure 2. Four
examples of randomly-generated synthetic prediction maps corresponding to a
given ground truth map. Candidate regions in synthetic prediction maps were
labeled in the same way as demonstrated in Figure 1, and radiomic features were
extracted in the same way as they were for candidate regions of prediction maps
obtained from the first-stage voxel-wise classifier.
Figure 1. (a)
Sample ground truth and prediction map generated from the first-stage
voxel-wise classifier (white = PCa). (b) Image dilation applied to maps, which facilitates identification of
candidate regions in the prediction map (four in this example) via
identification of connected voxels. (c) Labeling of candidate regions based on degree of overlap with voxels
in the ground truth map. Candidate regions are labeled HG-PCa only if ≥ 50% of
voxels within the region are labeled GS ≥ 4+3.