Radiation Induced Lung Damage Tissue Segmentation and Classification Framework
A Szmul1*, E Chandy1,4, C Veiga1, A Stavropoulou1, J Jacob1,2, D Landau3, C Hiley4, J McClelland1, (1) Centre for Medical Image Computing, University College London, London, UK, (2) Department of Respiratory Medicine, University College London, London, UK, (3) Guys & St Thomas NHS Foundation Trust, London, UK, (4) Cancer Institute, UCL, London, UK
Presentations
PO-GePV-M-5 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Radiation-induced lung damage (RILD) is a common side effect of radiotherapy. The ability to automatically segment, classify and quantify different types of parenchymal change is essential to uncover underlying patterns of RILD and their evolution over time.
Methods: The following five labels were introduced to classify parenchymal morphological appearance: A) default lung parenchyma; B) non-textured ground glass opacity; C) textured ground glass opacity; D) solid density with air-bronchograms; E) solid density. We utilized a two-stage ground truth data generation approach, akin to active learning. Manual segmentations were used to train an initial auto-segmentation method. These results were manually refined and used to train the final auto-segmentation algorithm.The auto-segmentation algorithm was an ensemble of six 2D-Unets with different loss functions and numbers of input channels.
Results: Our development dataset consisted of 40 cases, each with a pre-radiotherapy, 3, 6, 12, and 24 month follow-up CT scan (n=200 CTs). The method was assessed on a hold-out test dataset of 6 cases (n=30 CTs). The Dice Score Coefficient (DSC) achieved for each tissue class were: A) 0.99 and 0.98, B) 0.71 and 0.44, C) 0.56 and 0.26, D) 0.79 and 0.47, and E) 0.96 and 0.92, for development and test subsets respectively. The lowest values for the test subsets were caused by imaging artefacts or in regions where annotations were less certain. We also performed qualitative evaluation on the test dataset presenting manual and auto-segmentations to a blinded independent radiologist to rate them as ‘clinically acceptable’, ‘minor disagreement’ or ‘major disagreement’. The auto-segmentation ratings were similar to the manual segmentations, both having slightly over 80% of cases rated as ‘clinically acceptable’.
Conclusion: The proposed framework for auto-segmentation of different lung tissue classes produces acceptable results in the majority of cases and will be useful for future studies of RILD.
Funding Support, Disclosures, and Conflict of Interest: AS- EPSRC-funded UCL i4health CDT (EP/S021930/1). CV- RAE, Research Fellowship scheme (RF\201718\17140). IDEAL CRT trial- CRUK, C13530/A10424 and (C13530/A17007). AS, CV, EC, AS, JRM- charitable donation. JJ- Wellcome Trust Clinical Research Career Development Fellowship (209553/Z/17/Z), NIHR BRC at UCL JRM- CRUK Centres Network Accelerator Award Grant (A21993), ART-NET.
ePosters
Keywords
Lung, Segmentation, Tissue Characterization
Taxonomy
Contact Email