139. Implementation of a classifier based on a personalized atlas to validate contours and comparison of automatic segmentation algorithms in thoracic district: Atlas-based-segmentation vs. model-based-segmentation

Purpose Automatic segmentations to contour region of interest (ROIs) in radiotherapy is becoming an impending necessity to reduce physicians’ workload and inter-user variability. The aims were: (I) Comparison of automatic segmentations between atlas-based-segmentation (ABS) and model-based-segmentation (MBS); (II) Evaluation of ABS performances with different inputs; (III) Implementation of a classifier to assess segmentations quality. Methods and materials 42 patients (37 training, 5 validation), previously contoured within left partial breast irradiation protocol, were used for atlas construction. We focused on lungs, spinal cord, heart, contralateral breast and thyroid. Automatic segmentations were evaluated by DiceSimilarityCoefficient (DSC) and Hausdorff distance (AHD) computed in RayStation® by IronPython® scripts. ABS were launched, varying the number of structures and training patients. The first method contoured all the structures in one session searching the best matching with one training patient. The second one contoured each structure finding a one-by-one matching. To test the algorithms, 13 new patients were used. Moreover, TrueNegative cases were generated from original ROIs with translations/expansions of 0.2 cm, 0.3 cm, 0.5 cm and 0.8 cm. The classifier was implemented in Matlab®. Results ABS gained results for heart, spinal cord and breast similar with clinical gold standard. ABS performed auto-segmentation better than MBS for lungs (DSC: 0.97–0.97; AHD: 0.09–0.06 for left and right). The thyroid segmentation was not satisfactory (DSC Conclusion ABS using the customized atlas perform better than the MBS. ABS segmentations had performances comparable with clinical gold standard except for small ROI. ABS reduce physician contouring workload up to 75%, with a time gain of 20–30 min for patient.