Purpose: Auto-segmentation has become more widespread for contouring, saving time and decreasing inter-observer variability. In this study we analyzed the performance of both an atlas- and a deep learning-based auto-segmentation of patients undergoing breast radiation.
Methods: The atlas-based segmentation was built using 12 patients from our clinic in MIM v6.9.3. The model for the deep learning-based auto-segmentation (MIM Contour ProtégéAI) was provided by the vendor. CT scans of 12 unique patients were used for testing, and contours were manually delineated for four different structures (heart, left lung, right lung, and left breast) to be used as reference structures. The 12 test patients were then auto-segmented using the atlas and the deep learning-based methods. Three geometrical metrics were then used to quantitatively evaluate the auto-contours: the Dice similarity coefficient (DSC), the mean absolute distance (MAD), and the Hausdorff distance (HD).
Results: The mean DSC for the heart and the lungs were greater than 0.90 for both the atlas and deep learning techniques. For the breast, which was only available via atlas-based segmentation, the DSC value was 0.83±0.04 (mean ± standard deviation). The mean MAD for the lungs, heart, and breast were all less than 0.4cm for both techniques. The HD for atlas and deep learning techniques were 2.4±0.8cm and 3.1±0.6cm for the lungs, and 1.8±0.8 cm and 1.4±0.4cm for the heart, respectively. The atlas resulted in HD of 3.6±0.7cm for the breast. The largest discrepancies seen were in the lung and breast where the auto-contours tapered off at the inferior and superior edges.
Conclusion: Based on these results, we conclude that both auto-segmentation methods were sufficient for clinical use for heart and lungs, and the atlas-based segmentation will be implemented in our clinical workflow this year. The breast auto-segmentation will be used to facilitate automated beam setup for tangent fields.
Not Applicable / None Entered.