Conference proceeding
Leveraging High-Quality Research Data for Ischemic Stroke Lesion Segmentation on Clinical Data
2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pp.1-5
04/18/2023
DOI: 10.1109/ISBI53787.2023.10230775
Abstract
Modern Deep Learning models are often developed, tuned, and reported for extremely high-quality research data. Unfortunately, due to constraints of the clinical environments, much lower-quality data are routinely acquired, leading to a slow adaptation of segmentation models. Applying new deep learning models to the clinical domain requires special care in the training and deployment of those models. This work will guide model developers in creating image segmentation tools suitable for currently collected clinical data.Specifically, we identify strategies to utilize high-resolution research data for improving the performance of Deep Learning models for segmenting ischemic stroke lesions on clinical data. We analyze how improper manipulation of binary masks negatively affects model training, and how carefully constructed preprocessing steps allow improved training in the context of the widely accepted 3D-UNet architectures to achieve improved clinical performance. We demonstrate that naively adding high-quality research data to a training dataset does not ensure performance improvement However, we propose a method of manipulating research data to resemble clinical scans that significantly improves the model performance from 0.64 to 0.70 Dice coefficient. Our solution was also more robust to false positives and decreased the balanced averaged Hausdorff distance from 0.508 to 0.305 voxels. Additionally, we observed substantial performance improvement in segmenting small lesions, which are difficult to capture in a clinical setting.
Details
- Title: Subtitle
- Leveraging High-Quality Research Data for Ischemic Stroke Lesion Segmentation on Clinical Data
- Creators
- Michal Brzus - University of IowaAaron D. Boes - University of IowaJoel Bruss - University of IowaHans J. Johnson - University of Iowa
- Resource Type
- Conference proceeding
- Publication Details
- 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pp.1-5
- DOI
- 10.1109/ISBI53787.2023.10230775
- eISSN
- 1945-8452
- Publisher
- IEEE
- Language
- English
- Date published
- 04/18/2023
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering; Neurology; Electrical and Computer Engineering; Psychiatry; Stead Family Department of Pediatrics; Iowa Neuroscience Institute; The Iowa Institute for Biomedical Imaging; Neurology (Pediatrics); The Iowa Initiative for Artificial Intelligence; Iowa Informatics Initiative
- Record Identifier
- 9984461957402771
Metrics
62 Record Views