Session Item

Gross tumor volumes (GTV) of head and neck cancer (HNC) are difficult to identify in images, even with deep learning (DL), particularly when the primary tumor (GTV-T) and multiple nodal metastases (GTV-N) are present. Using only imaging for DL segmentation may lead to false segmentations. This study examines whether minimal user input, in which the oncologist only needs to single-click the lesions to be segmented, improves the detection ratio and segmentation performance of deep learning-based auto-segmentation for GTV-T and GTV-N of HNC.

We have included treatment planning CT, PET, and MRI (T1w mDixon and T2w) images for 567 HNC patients with a wide variety of tumor sites (larynx, pharynx, oral cavity, sinonasal, and salivary gland carcinomas). GTV-T and GTV-N clinical delineations were treated as two separate DL targets. The data was randomly split into training(n=375), validation(n=95), and test sets(n=97).

To simulate user input clicks, we generated a dot of random size between 5 and 10 mm³ at a random location inside each distinct target volume. We used this simulated user feed in conjunction with CT, PET, T1w, and T2w MRI scans as inputs to a 3D UNet. We compared the segmentation results to the UNet using only the scans as input.

We evaluated the detection ratio(%) on all the distinct GTV-Ts and GTV-Ns. The segmentation performance was evaluated using Dice Similarity Coefficient(Dice), Hausdorff distance 95%(HD95), mean surface distance(MSD), and Surface-Dice with 2mm tolerance. The voxel-based false discovery rate (FDR) and false negative rate (FNR) were used to measure false segmentation and compared using a Wilcoxon signed-rank test(p<0.05). FDR can be interpreted as an indicator of false positive segmentations, whereas FNR indicates false negative segmentations. For all metrics, the mean and 95% confidence interval (CI95, bootstrapping 10000 samples) were reported.

On the test set of 97 patients with 100 GTV-Ts and 177 GTV-Ns, after incorporating user feed, the detection ratio increased from GTV-T(95%)/GTV-N(79%) to GTV-T(99%)/GTV-N(99%). All segmentation metrics were improved, especially for GTV-N (Figure 1-A). The mean(CI95) of FDR for GTV-T/-N decreased from 0.27(0.23-0.30)/0.31(0.26-0.37) to 0.22(0.20-0.25)/0.16(0.14-0.19) with p<0.001. FNR of GTV-T was marginally affected by user input, reduced from 0.32(0.28-0.36) to 0.29(0.25-0.32) with p>0.05. Whereas FNR of GTV-N decreased significantly from 0.29(0.25-0.32) to 0.17(0.16-0.19) with p<0.001 (Figure 1-B). Two cases with significant improvement in GTV-N segmentation are plotted in Figure 2.

We improved HNC GTV DL auto-segmentation on a highly diverse data set,  where DL performance is typically poor. Single-click input per lesion reduced false negatives for GTV-N and false positives for GTV-T and GTV-N. This suggests that clinicians' prior knowledge could supplement medical scans, improving the detection ratio of GTV-N and the segmentation performance of GTV in DL auto-segmentation.