Synthetic cerebral blood vessel generator for training anatomically plausible deep learning models

Kenyon, Georgia; Lau, Stephan; Perperidis, Antonis; CHAPPELL, MICHAEL; Jenkinson, Mark

Abstract

Blood vessel networks, with their complex geometrical and topological characteristics, play a significant role in diagnosing and understanding various cerebrovascular diseases. Deep learning (DL) segmentation methods can aid in analysing these structures; however, models often produce anatomically implausible segmentations, overlooked by simple segmentation metrics. Extensive literature on cerebral vessel geometry rules, like branching patterns and vessel length-radius ratios, enable the creation of synthetic vessel label generators that can create data that adhere to or deviate from these rules. This data can be used to train DL networks, that score vessel label’s anatomical plausibility and implausibility. Trained networks can then be used to evaluate segmentation networks’ label outputs based on their anatomical plausibility, to go beyond commonly used, but mathematically simple, segmentation evaluation metrics. This work presents a novel synthetic cerebral vessel data generator, facilitating the generation of both anatomically plausible and implausible vasculature for the purpose of training DL models to assess the plausibility, or quality, of vessel segmentations in medical imaging.

Citation

Kenyon, G., Lau, S., Perperidis, A., CHAPPELL, M., & Jenkinson, M. (2024, July). Synthetic cerebral blood vessel generator for training anatomically plausible deep learning models. Presented at MIUA - Medical Image Understanding and Analysis - 2024, Manchester

• Presentation Conference Type: Edited Proceedings
• Conference Name: MIUA - Medical Image Understanding and Analysis - 2024
• Start Date: Jul 24, 2024
• End Date: Jul 26, 2024
• Acceptance Date: Jul 24, 2024
• Online Publication Date: Oct 7, 2024
• Publication Date: Oct 7, 2024
• Deposit Date: Oct 9, 2024
• Publicly Available Date: Oct 10, 2024
• Peer Reviewed: Not Peer Reviewed
• Book Title: Medical Image Understanding and Analysis, Manchester, UK
• ISBN: 9782832512449
• DOI: https://doi.org/10.3389/978-2-8325-1244-9
• Public URL: https://nottingham-repository.worktribe.com/output/40558593
• Publisher URL: https://www.frontiersin.org/books/Medical_Image_Understanding_and_Analysis/12759