Non contrast-enhanced imaging as a replacement for contrast-enhanced imaging for MRI automatic delineation of nasopharyngeal carcinoma

Convolutional neural networks (CNNs) show potential for delineating cancers on contrast-enhanced MRI. However, there is world-wide interest in reducing the administration of MRI contrast agents. We aim to determine if CNNs can automatically delineate primary nasopharyngeal carcinoma (NPC) using the non contrast-enhanced (NE) T2-weighted fat-suppressed T1-weighted (CE-T1W) sequence. We retrospectively analyzed primary tumors in 201 patients with NPC. Six patients were randomly sampled as the training-validation group to avoid over-fitting, and the remaining 195 patients underwent validation analysis. We trained and tested a well-established two-dimensional CNN, U-Net, for tumor delineation on CE-T1W and T2W-FS sequences. CNN-derived delineations on CE-T1W and T2W-FS were compared with manual delineation using the dice similarity coefficient (DSC) and average surface distance (ASD). Differences in DSC and ASD of CNN-derived delineations between CE-T1W and T2W-FS sequences were compared using the Wilcoxon rank test. CNN-derived primary tumor volumes (PTVs) on CE-T1W and T2W-FS were also compared with manual delineation using the Wilcoxon rank test. The CNNs tumor delineation performance on CE-T1W and T2W-FS showed no differences in DSC (0.71{+/-}0.09 vs. 0.71{+/-}0.09, p=0.50) and ASD (0.21{+/-}0.48cm vs. 0.17{+/-}0.19cm, p=0.34). The CNN-derived PTVs were larger than those from manual delineation on both CE-T1W (26.3{+/-}25.5cm3 vs. 23.5{+/-}26.6cm3, p<0.001) and T2W-FS (24.2{+/-}23.7cm3 vs. 23.2 {+/-} 26.2 cm3, p<0.001). In conclusion, CNN can automatically delineate primary NPC using the NE T2W-FS sequence which has the potential to be a substitute for the CE-T1W sequence. (T2W-FS) sequence and compare the performance with that from the contrast-enhanced