Sequential and iterative auto-segmentation of high risk clinical target volume for radiotherapy of nasopharyngeal carcinoma in planning CT images

Background: Accurate segmentation of tumor targets is critical for improving tumor control and normal tissue protection. In this study we proposed a novel sequential and iterative U-Net (SI-Net) convolutional deep learning approach to auto-segment the high risk clinical target volume (CTV) for radiotherapy of nasopharyngeal carcinoma (NPC). Methods: The SI-Net is a variant of U-Net architecture, which is a popular neural network architecture in biomedical image segmentation. The input of SI-Net model includes one CT image slice, the CTV contour on the current slice, and the next CT slice. The output is the predicted CTV contour on the next CT slice. We designed the SI-Net with left part to learn the volumetric features, and the right part to predict the contour of next image. Two predicted directions, one from inferior to superior (named as forward) and the other from superior to inferior (backward), were tested. The results were compared with that of standard U-Net model. Dice similarity coefficient (DSC), Jaccard index (JI), average surface distance (ASD) and Hausdorff distance (HD) were used to evaluate the segmentation performance. Three patients’ cases were selected to evaluated the inter-practitioner delineation variability. Results: The forward direction SI-Net model showed a 5% and 6% higher mean DSC and JI value than that of the U-Net model (0.84±0.04 vs 0.80±0.05 and 0.74±0.05 vs 0.69±0.05, p<0.001). The ASD and HD value of the SI-Net model also indicated a better performance (2.8±1.0mm vs 3.3±1.0 mm and 8.7±2.5 mm vs 9.7±2.7 mm, p<0.01). Both the DSC and JI of backward direction SI-Net model were still better than that of U-Net model (p<0.01), although the ASD and HD index were similar. Conclusions: The SI-Net model preserved the continuity between adjacent image slices, and thus improved the segmentation accuracy compared to the conventional U-Net model. This model will improve the efficiency and consistence of CTV contouring for NPC patients.