Click here to

Session: AI/ML Autoplanning, Autosegmentation, and Image Processing II [Return to Session]

Longitudinal Monitoring of Esophageal Expansion to Capture Acute Esophagitis Prior to Its Onset

H Veeraraghavan*, J Jiang, A Rimner, J Deasy, M Thor, Memorial Sloan Kettering Cancer Center, New York, NY


SU-F-BRB-9 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Ballroom B

Purpose: The goal was to develop deep learning (DL) based registration of during treatment longitudinal cone beam CTs (CBCTs) and identify early esophageal expansion in lung cancer radiotherapy (RT) that could be indicative of acute esophagitis (AE2).

Methods: A 3D DL recurrent registration network constructed with convolutional long short term memory networks was used to align CBCTs acquired once a week during the first two weeks of RT with the planning CT (pCT) of 45 stage II-III locally advanced non-small cell lung cancer (LA-NSCLC) patients treated with image-guided conventionally fractionated RT to a median of 60Gy. The maximum esophageal expansion (MEx%) was computed from the Jacobian determinant (J) of the deformation vector field and defined as the average voxel-wise tissue expansion in a 3 x 3 x 3 volume centered at a voxel with the highest J to measure local expansion (J -1 > 0) or shrinkage (J – 1 < 0). The change in MEx% from week1 to week2 (ΔMEx%) was compared between patients with and without AE2 using a two-sided Wilcoxon signed rank test.

Results: The DL method produced smooth deformations with tolerable folding of under 1%. It produced significantly more accurate (p < 0.001) esophagus segmentations than iterative diffeomorphic DIR (Dice similarity coefficient of 0.76 ± 0.12 vs. 0.72 ± 0.14, Hausdorff distance at 95th percentile of 3.83±2.83 mm vs. 4.96 ± 3.43 mm) and was fast (5secs vs. 10 mins). The ΔMEx% was significantly higher in patients with AE2 (p = 0.02; median 8.7%, inter-quartile range [IQR] of -8.0% to 44%) than those without AE2 (median -13%, IQR of -21% to 9.8%) indicating a higher degree of expansion among AE2 patients.

Conclusion: The proposed DL method successfully computed esophageal expansion/shrinkage metrics from longitudinal CBCTs and showed that patients with AE2 have a higher degree of expansion.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email