Session Item

Gross tumour volume (GTV) segmentation constitutes a time-consuming step in the radiation oncology workflow. It is essential for the treatment planning process or for radiomics analyses. In the case of soft tissue sarcomas, the presumed true GTV delineation can vary between different readers, as it may depend on different interpretations of contrast enhancement beyond the tumour bulk. The application of deep learning (DL)-based segmentation algorithms bears the potential to reduce the required work time while improving the inter-observer variance.

Based on T1-weighted contrast-enhanced MRI sequences, we used a training cohort (n = 157) to develop a 3DRes-UNet for GTV segmentation of soft-tissue sarcomas. The GTV was defined as the clearly delineable contrast enhanced tumour bulk, thus serving as the basis for radiomics analyses. Subsequent validation was performed using an independent test cohort (n = 87). Manual segmentations of a radiation oncologist were used as ground truths.

In a subgroup (n = 20), we performed a benchmark study: For each imaging study, two first-year residents (“early residents”, ERs) manually performed GTV segmentation and modified the DL-based GTVs. Similarly, two board-certified radiation oncologists (ROs) generated segmentations in a clinical approach including contrast enhancement extending beyond the tumour bulk. Further, the radiation oncologists evaluated the DL segmentations in a binary fashion according to clinical applicability. To reduce recall effect, the segmentation methods were randomly split into two sessions with a 4-week interval.

To analyse segmentation accuracy, GTVs were compared with ground truths by calculating the dice similarity coefficients (DSC) as well as the Hausdorff distance (HD). In addition, the time required for manual and DL-assisted segmentation was compared.

The algorithm segmentations achieved a median DSC of 0.88 (interquartile range (IQR): 0.10 and 0.07, respectively) in the entire test cohort and in the benchmark cohort. In the benchmark study, comparison of the manual and DL-based segmentations showed similar DSC and HD results within each physician group. Here, the median DSC was 0.92 (0.07) and 0.91 (0.06) for the ERs and 0.81 (0.11) and 0.83 (0.09) for the ROs, respectively. The median HD was 10.6 (12.3) and 8.4 (8.1) for the ERs and 18.7 (22.5) and 18.3 (22.7) for the ROs, respectively. The ROs rated the DL-based GTV segmentations as directly applicable for the further clinical planning process in 7/20 (35%) cases and 4/20 (20%) cases, respectively.

When comparing segmentation duration, there were no statistically significant differences between manual and DL-assisted implementation for either ERs or ROs.

The use of our DL-based algorithm provides suitable tumour volume segmentations. The generated GTVs can support manual segmentation, and will be further investigated for clinical application in the future.