ESTRO 2024 - Abstract Book

S3152

Physics - Autosegmentation

ESTRO 2024

3075

Digital Poster

The impact of slice thickness on an autosegmentation model for prostate MRI

Owen McLaughlin 1 , Alan R Hounsell 2,1 , Kevin M Prise 1 , Joe M O'Sullivan 3,1 , Suneil Jain 3,1 , Conor K McGarry 2,1

1 Queen's University Belfast, Patrick G. Johnston Centre for Cancer Research, Belfast, United Kingdom. 2 Belfast Health and Social Care Trust, Radiotherapy Physics, Northern Ireland Cancer Centre, Belfast, United Kingdom. 3 Belfast Health and Social Care Trust, Department of Clinical Oncology, Belfast, United Kingdom

Purpose/Objective:

Due to imaging protocols varying between institutes and image quality changing with time, it is necessary to establish the suitability of autosegmentation models applied to images with different properties. The purpose of this study was to test the impact of magnetic resonance imaging (MRI) slice thickness on an autosegmentation model.

Material/Methods:

An autosegmentation model was trained on an open-source dataset of 114 T2w axial MRIs with a slice thickness of 2 mm [1]. Images were converted to NIFTI file format and rotated for processing. Training was carried out using the MONAI Label framework [2], networks were trained for 150 epochs each. Each voxel in the image was segmented as the central gland, peripheral zone or image background. Fifteen images from this dataset were then each resampled to a range of slice thicknesses between 2 mm to 8 mm, keeping resampled pixel spacing constant ((1.5, 1.5) mm). In addition to examining prostate zones; the central gland and peripheral zone contours were summed to constitute the contoured prostate gland. A range of prostate gland volumes were included in the fifteen patient group ranging from 10.12 cc to 90.05 cc. Resulting autosegmentations were compared to the original manual segmentations using the 3DSlicer Segmentation Comparison extension (via Dice Similarity Coefficient (DSC), average Hausdorff Distance (HD) and absolute percentage difference in segmented volume to the reference volume). Statistical differences between the autosegmentation of resampled images to the autosegmentation of original images, of slice thickness 2 mm, were investigated using a Wilcoxon rank sum test, comparing their DSC, HD and volume differences.

Results: