Eclipse treatment planning system version 15

All patients were immobilized using the Body Pro‐Lok^TM platform (CIVCO system, Orange City, IA) in the supine position with their arms up. A Knee cushion was used to immobilize the knees. Patients were instructed to present for CT simulation and treatment with a comfortably full bladder and empty rectum. To ensure a relatively empty rectum, patients were instructed to use metamucil beginning 3‐day prior to simulation and daily throughout the course of treatment. A free‐breathing planning 3D‐CT scan was acquired on a GE Lightspeed 16 slice CT scanner (General Electric Medical Systems, Waukesha, WI) with 512 × 512 pixels at 2.5‐mm slice thickness in the axial helical mode. The planning 3D‐CT images were imported into Eclipse treatment planning system (TPS, Version 15.6, Varian Medical Systems, Palo Alto, CA) for contouring the entire prostate capsule as a clinical target volume (CTV) and the OAR. The planning target volume (PTV) was generated by adding a 3 mm symmetric margin around the prostate capsule per RTOG‐0938 recommendation.¹² The average PTV size was 117.0 ± 34.0 cc (range, 73.0 to 186.0 cc), corresponding to an average PTV diameter of 6.0 ± 0.6 cm (range, 5.2 to 7.1 cm). OAR contours included rectum, bladder, penile bulb, femoral heads, skin, and urethra per RTOG‐0938 requirement. In addition, small bowel and sigmoid were also contoured for dose reporting.

We investigated six target volumes: (1) GTV, (2) diminished GTV (GTV-2mm), (3) extended GTV (GTV+2mm), (4) PTV, (5) diminished PTV (PTV-2mm) and (6) extended PTV (PTV+2mm). GTV and PTV were delineated by clinical oncologists. the other four volumes (i.e., diminished GTV (GTV-2mm), extended GTV (GTV+2mm), diminished PTV (PTV-2mm) and extended PTV (PTV+2mm) were delineated by the study investigators using the Varian Eclipse treatment planning system version 15.6 (Varian, Palo Alto, CA, USA). All tumour volumes were checked by a certified medical dosimetrist.

Cases with multiple GTVs were combined into a single GTV for feature extraction by the Eclipse treatment planning system, version 15.6 (Varian, Palo Alto, CA, USA). The GTV was utilized for radiomic feature extraction performed by 3D slicer (v. 5.2.1, slicer.org) with Pyradiomics extension (Computational Imaging and Bioinformatics Lab, Harvard Medical). A total of 107 radiomic features were extracted from each sample (Table 1), which were imported into the machine learning algorithms. These radiomic features can be classified into seven groups, including shape, first-order feature, gray level co-occurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM) (see Table 1).

These patients were immobilized using the Body Pro‐Lok™ SBRT system (CIVCO, Orange City, IA) in the supine position with arms above the head. A free‐breathing helical CT scan was acquired on a SOMATOM.go CT simulator (Malvern, PA) with 512 × 512‐pixel image size and 1.25 mm slice thickness. Respiratory assessment and motion management included abdominal compression and a 4D‐CT scan via Varian RPM system (version 2.7). The 3D CT scan was brought into Eclipse Treatment Planning System (Version 15.6, Varian Medical Systems, Palo Alto, CA). With a 4D‐CT scan, an internal target volume (ITV) was contoured based on the registered 4D‐CT reconstructed maximum intensity projection (MIP) images. The planning target volumes (PTVs) were created by expanding a uniform margin of 5 mm from each ITV. The two targets were labeled arbitrarily as PTV1 and PTV2 by the treating physicians. All planning was completed on the free‐breathing untag CT images and Hounsfield units within the PTV were maintained per the planning CT dataset following our in‐house SBRT protocol. Critical structures were delineated including lungs (right, left and combined), spinal cord, heart/pericardium, trachea and bronchus tree, esophagus, skin and ribs (right, left and combined).