br Three patients presented with absolute offsets of nine
Three patients presented with absolute offsets of nine cm or greater at baseline or week six, which is an offset consistent with 5% ADC bias. The slice offset differences between time-points for these patients were 29, 52, and 68 mm. Exclusion of these three patients had minimal effect on the sets of extracted radiomics features.
ADC maps with and without b = 0 s/mm2 DWI were analyzed be-cause not all scanners and platforms can generate PIRADS v-2 com-pliant ADC maps with lowest b-value of 50–100 s/mm2, and because features which are consistently significant may be more robust response metrics. The ADC maps including b = 0 s/mm2 were derived in-line, Physics and Imaging in Radiation Oncology 9 (2019) 1–6
and the PIRADS v-2 compliant ADC maps excluding b = 0 s/mm2 dif-fusion-weighted images were derived using Matlab (Mathworks, Natick, MA), via weighted least squares regression to: Log(S(b)/S
(bmin)) = −(b− bmin)* ADC + c, where bmin denotes the lowest b-value used for the regression and c is an arbitrary baseline. Weightings were proportional to the inverse of the signal variance. ADC accuracy and signal-to-noise adequacy for this protocol at b = 1000 s/mm2 has been confirmed .
2.4. Tumor identification
Tumors were identified according to PIRADS v-2 parameters on treatment planning system (Pinnacle). Delineation was performed manually by expert radiation oncologists (CM, PC) to encompass all suspicious voxels on multiparametric MRI. In cases with multiple le-sions, boost was applied to all lesions.
2.5. Deformable registration
Deformable registration between baseline and week six image sets and segmented volumes was performed to increase robustness to pos-sible intra-scan motion, variations in prostate gland volume, shape, and orientation between imaging time-points [17,18], and loss of intra-prostatic image Lipo2000 post-EBRT . Prostate boundaries and at least three common points were contoured on baseline and week six T2-weighted images, using MIPAV software (NIH, Bethesda, MD). The points provided an initial rigid alignment, and MORFEUS , a bio-mechanical-based deformable registration technique, computed dis-placement from baseline to the week six T2w GTV guided by the prostate surface. The deformable registration accuracy was measured by target registration signed error (TRE), calculated from the dis-placements between the observed and the MORFEUS-predicted point coordinates.
GTVs were drawn on baseline T2-weighted images, guided by characteristic tumor hypointensity in pre-treatment ADC maps, and then deformably registered to week six T2-weighted images. The baseline and week six GTVs were then applied to their corresponding ADC maps using MIPAV, and manually corrected as deemed necessary between ADC and T2-weighted images, to account for routine ADC distortion and inter-acquisition motion . The extent of manual correction was quantified by calculation of the shift in Centre-of-Mass of each GTV using MIPAV.
Fig. 1 presents representative T2-weighted images, ADC maps, and GTV at each time point. Across all sets, TRE was calculated from 185 points corresponding to fiducial markers and/or natural landmarks. The average and standard deviation TRE was 0.7 ± 3.8 mm, 0.2 ± 2.9 mm, and 0.1 ± 6.9 mm in the LR, AP and SI directions re-spectively. The average magnitude of error vector was 4 ± 7 mm. Manual corrections of GTVs applied to ADC images from T2-weighted images were performed in 35 GTVs at baseline, and in 35 GTVs at week six. The mean and standard deviation vector shifts in GTV centres-of-mass were 2.3 ± 1.6 mm. Twenty vector shifts were greater than 3 mm, and eleven vector shifts were greater than 4 mm. In some cases, these GTV shifts were corrections from partially outside of the prostate gland or between zonal regions.
2.6. Radiomics feature extraction
Radiomics analysis used the open-source pyradiomics package , customized for feature extraction from GTV and prostate ROIs applied to baseline and week six ADC maps, and corresponding T2-weighted images. A total 101 radiomics features were extracted, which comprised the available pyradiomics feature set, excluding general image-speci-fying features which were not meaningful for signal characterization (e.g. BoundingBox, EnabledImageTypes; GeneralSettings; ImageHash; ImageSpacing; MaskHash; Version).
2.7. Statistical analysis
Feature values were compared using two-tailed Student’s t-test in Matlab, first between SIB and HDRB cohorts at baseline and week 6, and then between baseline and week 6 time-points using the pooled SIB and HDRB patient cohort. Correlations between the prostate and GTV feature values at baseline, week six, and their difference, were in-vestigated using Pearson’s correlation coefficient. These correlations were performed as part of a hierarchical clustering analysis (cluster-gram function, Matlab). In total, each feature was compared 603 times, comprised of two t-tests between SIB versus HDRB cohort values at baseline and at week 6; one t-test between baseline and week 6 for the pooled cohort values changes; and 600 comparisons from the hier-archical clustering analysis. For example, each baseline GTV ADC fea-ture was correlated with 100 baseline, week six, and difference features for both ADC and T2w image sets. The corresponding p-value for sig-nificance is 0.00008 based on Bonferroni correction.