Abstract
Rule-based fiber models, such as the transmural helix angle (HA) rule of Streeter et al., are widely used as a practical substitute for cardiac diffusion tensor imaging (cDTI) when assigning fiber orientation to left ventricular (LV) geometries, since they require only the wall geometry and no subject-specific imaging. However, these rules assign the same transmural HA profile at every circumferential and longitudinal position, regardless of local anatomy - a limitation that real cDTI data could correct, though it remains costly to acquire. We investigate whether finetuning on a limited number of real cDTI hearts can capture the regional variations that this transmurally homogeneous rule cannot represent. LV geometries were expressed in a topology-agnostic ventricular coordinate system (ϕ, ρ, θ) (circumferential angle, transmural depth, longitudinal position), in which a 3D U-Net was trained to map the LV wall mask alone to a transmural HA field. The network was first pre-trained on Streeter-rule HA maps, which can be generated for any wall geometry without real imaging, and then finetuned on real GT HA maps derived from cDTI eigenvector decomposition in a cohort of ovine and porcine hearts. Both the Streeter rule and the proposed pretrained-then- finetuned model were evaluated against real GT HA. On a held-out test set, the finetuned model achieved a mean angular error of 27.84◦ (median 20.78◦ ), compared to 45.40◦ (median 36.71◦ ) for the Streeter rule alone - a reduction of over 17◦ on average. This two-step approach - start from a rule-based estimate, then correct it with a limited number of real mea- surements - gives more accurate fiber orientation than the rule-based estimate alone, and could make subject-specific fiber modeling practical even when cDTI data is scarce.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/STACOM2026_045.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Adrien_Junior_Tchuem_Tchuente1
BibTex
@InProceedings{TchAdr_Deep_MICCAISAT2026,
author = { Tchuem Tchuente, Adrien Junior AND Cedilnik, Nicolas AND Rodríguez Padilla, Jairo AND Mojica, Mia AND Magat, Julie AND Pop, Mihaela AND Ozenne, Valery AND Cochet, Hubert AND Sermesant, Maxime},
title = { { Deep Learning Helix Angle Estimation via Rule-based Pre-training and cDTI Fine-tuning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17267},
month = {pending},
page = {pending}
}