Purpose: Deep learning (DL)-based auto-segmentation models has demonstrated state-of-art contour quality. There are multiple DL models to segment various treatment sites developed, validated and published in our group. This study aims to build a clinical pipeline to deploy the multiple DL-based models and automate the organs at risk contouring for clinical treatment planning.
Methods: To fully integrate the auto-segmentation tool with treatment planning workflow and minimize human intervention, we utilized the application programming interface (API) in our treatment-planning-system (TPS) to bridge up TPS with the server hosting the DL models. There are three models, head-and-neck, prostate, gynecology (GYN), implanted in the clinical pipeline and evaluated in this work. We deployed the models in two phases: (1) performance test in clinical pipeline to evaluate the quality and workflow readiness (2) performance monitoring after the clinical deployment. The contour quality is evaluated by comparing the auto-segmented contours and the physician reviewed contours and quantified by Dice Similarity Coefficient(DSC), Hausdorff Distance(95%)(HD95), and the frequency of the surface distance(SD) within 2 mm.
Results: Head and neck model was the first one deployed in the clinic and has been used in clinic for over 400 cases. The model exhibited in DSC 0.86 ±0.15 (Range: 0.76-0.96), HD95 4.07 (Range: 1.78-7.75) and SD 0.84±0.18 (Range: 0.71-0.97). The compliance rate of physicians completing contours on time improved from 67% to 93%. The prostate model segments bladder, rectum, prostate and femoral heads. We tested 40 cases in clinical pipeline in phase-1 and observed DSC 0.86 ±0.12 (Range: 0.80-0.93), HD95 7.73 (Range: 4.81-10.16) and SD 0.57±0.20 (Range: 0.41-0.62). Despite the satisfactory contour quality in phase-1 test, GYN model is under re-training to eliminate manual define range of segmentation based on the clinical feedback.
Conclusion: Clinical implementation of DL-based auto-segmentation improved the quality and efficiency of planning workflow.
Not Applicable / None Entered.