Gleason Probability Maps: A Radiomics Tool for Mapping Prostate Cancer Likelihood in MRI Space

Abstract

Prostate cancer is the most common noncutaneous cancer in men in the United States. The current paradigm for screening and diagnosis is imperfect, with relatively low specificity, high cost, and high morbidity. This study aims to generate new image contrasts by learning a distribution of unique image signatures associated with prostate cancer. In total, 48 patients were prospectively recruited for this institutional review board-approved study. Patients underwent multiparametric magnetic resonance imaging 2 weeks before surgery. Postsurgical tissues were annotated by a pathologist and aligned to the in vivo imaging. Radiomic profiles were generated by linearly combining 4 image contrasts (T2, apparent diffusion coefficient [ADC] 0-1000, ADC 50-2000, and dynamic contrast-enhanced) segmented using global thresholds. The distribution of radiomic profiles in high-grade cancer, low-grade cancer, and normal tissues was recorded, and the generated probability values were applied to a naive test set. The resulting Gleason probability maps were stable regardless of training cohort, functioned independent of prostate zone, and outperformed conventional clinical imaging (area under the curve [AUC] = 0.79). Extensive overlap was seen in the most common image signatures associated with high- and low-grade cancer, indicating that low- and high-grade tumors present similarly on conventional imaging.

Keywords: prostate Cancer; rad-path; radio-pathomics; radiomics.

Figure 1.

Generation of Radiomic profiles. Left: The 4 contrasts included in this study and the resulting segmentations created using Otsu's method. Right: 81 unique image characteristics created by linearly combining the segmented image contrasts. Each voxel receives a 4-digit code representative of the segmented image contrasts. Code 1133 indicates dark ADCshort, dark ADClong, bright T2, and bright DCE.

Figure 2.

Generation of Gleason probability maps from a training data set. Top: Hematoxylin and eosin stained whole mount histology and the corresponding pathologist annotations and T2 slice. Middle: Radiomic profiles are masked by the pathologist annotations and the distribution of the radiomic profiles. Bottom: The distribution of radiomic profiles within high grade, low grade, and benign regions are analyzed over 32 patients. The resulting probability values are applied to the radiomic profiling images on naïve data to create Gleason probability maps.

Figure 3.

Top: Radiomic profiles generated with global thresholds calculated on 3 different training sets, applied to a patient not otherwise included in the analysis. Bottom: Overlap map: yellow pixels have an identical image signature across all cohorts, red and yellow pixels have identical image signatures in 2 cohorts. Blank pixels have no overlap.

Figure 4.

Top: T2-weighted image and deep annotation overlaid on the same slice. A grade 4 cribriform tumor is shown in yellow. Bottom: Gleason probability maps created with and without the inclusion of zone information in the training data set. The images are nearly identical and the tumor is clearly visible.

Figure 5.

Receiver operator characteristic (ROC) evaluating the performance of the 4 raw image contrasts compared to Gleason probability maps (cancer probability). ADC 50-2000 = 0.78, ADC 0-1000 = 0.77, DCE = 0.65, T2 = 0.57, Gleason probability map = 0.79.

Figure 6.

Gleason probability maps. Top: True-positive cases. High-grade tumors are shown on the deep annotation in pink (cribriform) and yellow (not cribriform). Low-grade tumors are shown in green. Images are scaled to reflect the maximum probability in the training data set. Bottom: True negative. The displayed slide has only benign atrophy, and thus, no hot spots occur in the Gleason probability maps.