Feasibility of Using a Deep Learning Auto-Segmentation Software Trained with Planning CT for Iterative CBCT Based Online Adaptive Prostate Treatment

Abstract

<h3>Purpose/Objective(s)</h3> Cone-beam CT (CBCT) based organs-at-risk (OARs) delineation is the prerequisite of online adaptive therapy (ART) which can be time-consuming and inefficient. Auto-segmentation on CBCT would reduce the extra clinical resources required, however, it is labor-intensive to retrain the deep-learning auto-segmentation (DLAS) with CBCT contours that need to be labeled. Over standard CBCT, iterative CBCT (iCBCT) yields higher quality images with reduced noise and artifacts. This study aimed to comprehensively evaluate the feasibility of DLAS software trained with planning CT for iCBCT based online adaptive prostate treatment. <h3>Materials/Methods</h3> Total 25 male pelvis iCBCTs from corresponding prostate patients were selected for this study. An automated treatment planning process was established to simulate the online ART procedure by combining CT-based commercial DLAS software (i.e., trained with planning CT) and knowledge-based treatment planning used to eliminate human bias. Prostate and surrounding critical structures (i.e., bladder, rectum, and femoral heads) were delineated on iCBCT by a CT-based DLAS and by manual modification from corresponding planning CT registration. The geometrics metrics of OARs were computed between DLAS contours and manual contours. The prostate, considered as the gross tumor volume (GTV), was manually modified from DLAS to pursue accurate target dose coverage. For each iCBCT, two VMAT plans of 70 Gy with two full arcs were generated using the manual contour sets and DLAS contour sets respectively, which share the same modified prostate contour. Both plans were normalized to 100% of the prescription dose to cover 98% of the planning target volume (PTV) derived from GTV. The dose distributions from two plans were evaluated on the manual structure sets. The clinical appropriateness was evaluated by assessing D15(Gy), D25 (Gy), D35(Gy), and D50 (Gy) of critical structures following the RTOG-0815. The time required for the automated treatment planning process was recorded. <h3>Results</h3> Average dice agreement for bladder, rectum, femoral head_L and femoral head_R were 0.87 ± 0.10, 0.82 ± 0.08, 0.91 ± 0.12, and 0.93 ±0.07 respectively. DLAS generated a statistically significant of 0.39 Gy greater on bladder D₂₅ than its counterpart. No statistically significant differences were found in other OARs dosimetric metrics. All unmodified OARs satisfied the dose constraints of RTOG-0815 even with some artifact cases involved. The average time needed for the automated treatment planning process simulating ART was 11.85 minutes including DLAS generated time (0.69 minutes), GTV modified time (0.9 minutes), and plan generated time on CPU (10.26 minutes). <h3>Conclusion</h3> The proposed DLAS trained with planning CT is a promising contouring solution for iCBCT-based intact prostate online ART in the clinic with labor shortage. Without modification needed, it can generate clinically acceptable OARs segmentation on iCBCT images within a limited time.