Main Session
Sep 28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology

3050 - Deep Learning Auto-Segmentation for Quantification of Inter-Fraction GI Organ Positional Uncertainty on Serial Abdominal MRI

10:45am - 12:00pm ET
Poster Hall - Exhibit Hall A
Screen: 27
POSTER

Presenter(s)

Jeril Lasington, MD, MS, MBBS - New York Medical College at St.Mary's and St. Clare's, Denville, NJ

J. Lasington1, L. S. Mathew Lasington2, S. Umamaheshwaran3, and J. Thomas4; 1New York Medical College at St. Mary's and St. Clare's, Denville, NJ, 2Rutger's University, East Hanover, NJ, 3Personio, Manhattan, NY, 4Christian Medical College, Vellore, India

Purpose/Objective(s): To develop a deep learning auto-segmentation model for gastrointestinal organs at risk on abdominal MRI and to quantify the magnitude of inter-fraction organ positional uncertainty across serial scans in patients undergoing radiotherapy.

Materials/Methods: A U-Net with EfficientNet-B4 encoder was trained on the UW-Madison GI Tract dataset (107 patients, 64,368 MRI slices) with expert contours of stomach, small bowel, and large bowel. Segmentation was evaluated using Dice similarity coefficient (DSC) and 95th-percentile Hausdorff distance (HD95) on a held-out test set (193 scans). Centroid localization accuracy was assessed against expert contours on the same scans. Inter-fraction displacement was quantified using expert-contour 3D centroids across 96 patients with serial imaging (median 5 timepoints/patient), with directional decomposition (AP, LR, CC). Van Herk margins (M=2.5S+0.7s) were calculated and compared against published planning margins (RTOG 10mm, MR-Linac 3mm).

Results: Auto-segmentation achieved DSC of 0.91 (stomach), 0.88 (large bowel), 0.78 (small bowel), with centroid localization error of 4.4-11.4mm vs expert contours. Expert-contour inter-fraction 3D displacement was substantial: 41.0±26.9mm (large bowel), 37.4±27.0mm (stomach), 36.6±21.1mm (small bowel). At short intervals (=7 days), displacement remained 28-34mm. The CC axis contributed the largest component (25.9-30.5mm). 98% of patients (92% of fractions) exceeded 10mm displacement. Van Herk margins were 57-75mm, far exceeding conventional 10mm CTV-PTV margins, which were exceeded in 96% of fractions.

Conclusion: Deep learning auto-segmentation achieved clinically acceptable GI organ accuracy with centroid localization within 4-11mm of expert contours. Inter-fraction positional uncertainty was substantial, exceeding standard margins in nearly all patients regardless of time interval. These findings support routine daily online adaptive replanning in MR-guided abdominal radiotherapy, as conventional margins are insufficient for observed organ displacement.

Segmentation Accuracy, Inter-Fraction Displacement, and Calculated Margins

DSC - Dice Similarity Coefficient,HD95 -95th-percentile Hausdorff Distance,Centroid Err - 3D distance between centers of auto-segmented organ and expert-contour,3D Disp-3D displacement,AP -Anterior-Posterior,LR -Left-Right, CC -Cranio-Caudal ,=7d Disp- Displacement measured between scans =7 days apart, >10mm -Percentage of patients/fractions exceeding 10mm displacement threshold,vH M - Van Herk margin

Metric

Stomach

Small Bowel

Large Bowel

AP (mm)

LR (mm)

CC (mm)

=7d Disp

>10mm

vH M

DSC

0.91

0.78

0.88

-

-

-

-

-

-

HD95 (mm)

11.4

25.4

24.3

-

-

-

-

-

-

Centroid Err

4.4±3.7

10.7±8.9

11.4±13.8

-

-

-

-

-

-

3D Disp (mm)

37.4±27.0

36.6±21.1

41.0±26.9

-

-

-

-

-

-

Directional

-

-

-

14.0

20.8

25.9

-

-

-

=7d Disp

28.1±23.3

28.6±20.4

33.7±26.2

-

-

-

-

92%

-

S (mm)

22.1

19.2

25.1

-

-

-

-

-

-

s (mm)

19.0

12.7

17.7

-

-

-

-

-

-

vH M

68.5

56.9

75.1

36.2

54.6

81.0

73-84

98%pts

>>10