Patient-Specific Daily Updated Deep Learning Auto-Segmentation for MRI-Guided Adaptive Radiotherapy

Abstract

<h3>Purpose/Objective(s)</h3> MR-guided adaptive radiotherapy (MRgART) requires accurate and efficient daily image segmentation for plan adaption. Deep learning (DL) technique has shown great potential on medical image segmentation, but still has limited success on MRIs, partially due to the large variations for MRIs from different scanners, protocols, and patients. The longitudinal MRIs and contours during MRgART contain the anatomy variation for the specific patient. However, it has been ignored in current auto-segmentation solutions. This study proposed a patient-specific DL auto-segmentation strategy using patient's previous images and contours to update a prototype model, to improve segmentation accuracy and efficiency for MRgART. <h3>Materials/Methods</h3> For each patient, a prototype model was trained based on SegResNet using the first set of MRI scan and the corresponding contours as inputs. The patient-specific prototype model was updated with each newly added fractional MRI/contours and then used to predict contours for next MRI of the specific patient. During model training, a mutant of the prototype tried to fit the patient's previous data under consistency constraints, which is to limit the difference between the predictions for the latest MRI (volume, length, centroid) within a reasonable range. Prototype weights were partially inherited from the mutant using stochastic weight averaging to get the final model. Six abdominal or pelvic patients (each with 8 longitudinal MRIs) underwent MRgART on a 1.5T MRI-Linac were utilized. A total of 16 abdominal and pelvic organ-at-risks (OARs) and the tumors were manually contoured by three senior radiation oncologists and used as ground truth. The model performance was compared with deformable image registration (DIR) and frozen DL model using Dice Similarity Coefficient (DSC). The contouring time including model inference and contour reviewing and editing time until physician's approval was recorded and compared with fully manual contouring. <h3>Results</h3> The proposed daily updated model achieved superior performance with average DSC of 0.907±0.117, as compared to DIR (0.874±0.109) and frozen DL models (0.889±0.121). As for tumors, the proposed method achieved an average DSC of 0.922±0.097, which is significantly higher than those from DIR (0.723±0.183) and frozen DL (0.591±0.232). The DSC increased by average 3.1% with one additional prior images/contours compared with the prototype model. The contouring time is significant shorter using proposed method (average 6 minutes) than the manual process (20∼40 minutes). <h3>Conclusion</h3> The proposed patient-specific DL auto segmentation method can significantly improve the efficiency and accuracy of contour generation on longitudinal MRIs. It can be integrated into the daily adaption workflow, accelerating the contouring process, thereby facilitating routine practice of MRgART.