Session Item

We have initiated the ESO-SPARE phase III trial where patients with metastatic spinal cord compression are randomized to either esophagus sparing- or standard IMRT with patient reported esophagus toxicity as primary endpoint. Another aim is to reduce time to treatment by using artificial intelligence (AI) software to automate the planning process.  

First step is to use AI automatic segmentation of the esophagus. Manual delineations are prone to interobserver variation and we hypothesize that the variation between automatic segmentation and manual delineation has the same magnitude as interobserver variation. Furthermore, we hypothesize that the dosimetric consequence of using the automatic segmented esophagus for plan optimization is minimal compared to a manual delineated esophagus.

The planning CTs of ten consecutive patients referred for palliative radiotherapy for metastatic spinal cord compression (MSCC) in the thoracic spine were included in this analysis. The scans were acquired with a 2-mm slice thickness on a Siemens SOMATOM go.Open Pro with syngo CT VA30A software installed (Siemens Healthineers ™). Automatic segmentation of the esophagus (eso-AI) was done using the scanner software. Two radiation oncologists retrospectively delineated the esophagus in full length (eso1/eso2) - blinded to the eso-AI and each other. The esophagus’s lengths were adjusted to start and stop in the same slices. VMAT plans with two 360 degree-arcs were created for each of the ten patients with optimization on each of the three eso-volumes. Prescribed dose was 25 Gy/5 fractions. The esophagus constraint (D_(0.027cc) < 8.5 Gy) was prioritized higher than the PTV constraint (V_90% > 95%). Delineation and treatment planning were performed in eclipse (Varian medical systems ™). DICE similarity scores were calculated for eso-AI/eso1, eso-AI/eso2 and eso1/eso2. Dose coverage was compared for the three plans optimized on each patient. Friedmans test are used for statistical analysis and results are reported as median (range).

The mean DICE similarity scores were significantly smaller between eso-AI and the two observers (0.69-0.84 and 0.70-0.83) than between observers (0.79-0.88) (p<.0005). The eso-AI had a significantly smaller volume than both eso-1 and -2 (p=0.02). In one case, the eso-AI was not segmented in all relevant slices, Figure 1. There was no difference between target coverage, measured as PTV V_90%, GTV V_95% and GTV D_mean, in plans optimized on the different esophagus volumes. When the eso-AI was used for optimization it led to overdosing of eso1 and eso2, Table 1.

The automatic segmented esophagus was smaller than the manual delineated esophagi and even failed to segment the entire esophagus in a single case.  When used for optimization it led to violation of the dose constraint for the manual delineated esophagi. However, the clinical impact is unknown. The automatic segmented esophagus can not be recommended for clinical use without manual correction.