2.3 Imaging segmentation
We defined three regions of interest (ROIs) in CT images, including gross target volume (GTV), pelvis, and sacral vertebrae (Figure 1). GTV consisting of the visible primary tumor and positive pelvic lymph node was previously defined for radiotherapy planning based on both CT and MRI (contrast-enhanced T1-weighted images, T2-weighted images, DWI images) by the consensus of experienced radiation oncologists and radiologists. The size and morphology of all visible lymph nodes were inspected on CT or MRI images by two oncologists (10 years of experience) and reviewed by one radiologist (30 years of experience). A positive lymph node was a rounded soft-tissue structure with a short-axis diameter > 10mm or with central necrosis [27, 28]. The pelvis and sacral vertebrae were contoured independently by two radiologists (10 years of experience) and confirmed by a radiologist (30 years of experience). We used radiomic features of GTV to predict treatment outcomes, while those of pelvis and sacral vertebrae to predict hematologic toxicities. In addition, one radiologist (7 years of experience) independently contoured GTVs for 30 of 257 CT cases to test the inter-rater reliability of radiomic features.
2.4 Assessment of treatment outcomes and hematologic toxicities
The short-term tumor response was assessed based on pelvic MRI examination 3 months after radiotherapy according to the Response Evaluation Criteria In Solid Tumors v. 1.1 [29]: (1) CR represents the disappearance of all cervical lesions; (2) partial response (PR) represents at least a 30% decrease in the longest tumor diameter; (3) progressive disease (PD) represents at least a ≥ 20% increase in the longest diameter of tumor; and (4) stable disease (SD) represents neither sufficient decrease to qualify for PR nor sufficient increase in longest diameter for PD. OS was defined as the time from the date of diagnosis until death or the last follow-up. Hematologic toxicities during chemoradiation were assessed according to the National Cancer Institute Common Terminology Criteria for Adverse Events CTCAE 3.0 [29].
We investigated four clinical endpoints. (1) we categorized patients into the CR group and the non-CR group to predict tumor regression. (2) patients were divided into two groups by the cutoff OS time of 5 years. Binary classification models were used to predict whether patients survived ≥ 5 years. (3) we predicted whether patients suffered from severe leucopenia (grade ≥ 3) and (4) hypohemoglobin (grade > 0) that could reflect treatment tolerance and influence outcome. Table 2 shows the number of patients stratified by the clinical endpoints.
Additionally, we used one-class learning algorithm to identify patients at high risk of treatment failures (SD and PD) due to the small sample size (5 SD and 4 PD of 257 patients). Methods and results are in supplementary materials.