Radiomics feature stability and distinction power in organic low-contrast phantoms for novel CBCT imaging.

Purpose: Due to the prevalence of daily cone-beam computed tomography (CBCT) imaging in radiation therapy, radiomics analysis has great potential to detect early radiation induced tissue changes. Clinical applications of radiomics using CBCT imaging have been hindered by lack of stability in radiomics features and comparably poor image quality. Novel CBCT imaging devices promise improved quality comparable to those of fan-beam CTs. Using organic low-contrast phantoms and patient data, this study aims to assess the potential of novel CBCT imaging for radiomics analysis by testing the robustness and utility of extracted features.

Methods: CBCT scans (Hypersight, Varian Medical Systems, Palo Alto, CA) of three groups of low-contrast organic phantoms (16 apples, 16 oranges and 16 onions) were acquired using four different clinical scan-presets (head: 100 kV, 88 mAs; head low dose: 100 kV, 29 mAs; pelvis: 125 kV, 470 mAs; breast: 125 kV, 29 mAs). Each scan was repeated without change (re-test). The phantoms were then moved and repositioned to the same position for a third scan (reposition-test). Lastly, the phantoms were rotated by 90° clockwise and scanned again (90°-test). 107 radiomics features were extracted including shape, first- and second-order features. The concordance correlation coefficient (CCC) of these features between the initial scan, re-test, reposition-test and 90°-test was determined, respectively. Features with a CCC greater than 0.90 were deemed stable. A Boruta random-forest analysis was conducted for each scan-preset and the features were ranked by their importance (z-score) in distinguishing between the phantom groups. When a feature presented a z-score higher than 2.58 (p = 0.01) it was deemed viable. Correlation cluster plots were generated to visualize redundancies. Finally, the presented analysis was repeated on clinical image data of 16 primary prostate cancer patients. Two CBCTs were acquired (pelvis preset) before the initial treatment with 24.0 (20.3, 28.6) minutes between them. Prostate, rectum and bladder were segmented and all features were extracted per organ. The CCC was calculated for these image pairs per organ analogous to re-test. A Boruta random-forest analysis was used to identify features that are viable in distinguishing the three organs.

Results: The average fraction of stable features over all phantom groups and scan-presets in re-test was 100.0 %, 98.1 % and 98.4 % for shape, first-order and second-order features, respectively. In reposition-test 97.0 %, 90.3 % and 96.2 % were stable and in 90°-test 86.3 %, 75.9 % and 65.8 %. Feature stability rate was comparable between different scan-presets with the highest stability rate being head (89.1 %) followed by breast (88.0 %), head low dose (87.7 %) and pelvis (87.3 %). Boruta random-forest analysis yielded the following features to be most relevant in distinguishing the phantom groups: gray level co-occurrence matrix (GLCM) "inverse variance" in head-preset (6.01 ± 0.54) and pelvis (6.00 ± 0.51). GLCM "IMC1" was the highest scoring feature for head low dose (5.49 ± 0.54) and breast (5.47 ± 0.48). The total number of features viable in discriminating the phantom groups were 46 in breast-preset, 43 in head-preset, 44 in head low dose-preset and 43 in pelvis-preset. The fraction of stable features in the clinical example were 63 %, 15.0 % and 15.0 % in prostate, rectum and bladder, respectively. Boruta random-forest analysis yielded 36 viable features for discriminating the organs.

Conclusion: Radiomics analysis using novel CBCT imaging yields a high rate of stable features in re-test and reposition-test for all scan-presets and phantom groups. Feature stability rate is largely independent from the chosen scan-preset with only 1.8 % difference between best and worst stability rate. The statistical analysis yielded many texture-based features that were viable for discriminating the phantom groups. The clinical data analysis also produced many viable features for discriminating the three low-contrast organs. These findings demonstrate that the image quality is sufficient for radiomics analysis using clinical low-contrast data.