Abstract
Deep-inspiration breath-hold computed tomography (DIBH CT) is used in left breast cancer radiotherapy to reduce cardiac exposure, while free-breathing CT (FB CT) remains the more routine acquisition setting. FB annotations can provide useful supervision for DIBH segmentation, but direct transfer is limited because the clinical target volume (CTV) and organs-at-risk (OARs) change position and shape between respiratory states. We study label-efficient DIBH segmentation and propose a structure-state conditioned teacher–student framework for transferring FB knowledge to DIBH. An FB teacher is first trained from annotated FB CT. Paired FB/DIBH scans are then used to learn a contour-specific respiratory transition that warps FB teacher logits into the DIBH state and applies a residual correction before distillation. This design allows each CTV or OAR contour to use a different part of the FB-to-DIBH transition, rather than forcing all structures to share one mapping. Reliability weighting further reduces the influence of uncertain transferred labels near displaced boundaries. We evaluate the method on 148 patients with paired FB and DIBH CT scans for left breast cancer radiotherapy, including CTV and five OARs: heart, left lung, right lung, right breast, and left anterior descending coronary artery (LAD). Compared with FB-only, DIBH-only, Joint FB+DIBH, Registration propagation, Standard knowledge distillation, and Shared-transition baselines, the proposed method achieves the best mean DSC and HD95. It reaches 91.2% mean DSC and 5.1 mm mean HD95 with all DIBH labels, and maintains larger gains when DIBH labels are reduced.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Deep_Brea3th_036.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/forum?id=LFjfzxMwcH
BibTex
@InProceedings{DaoBil_StructureState_MICCAISAT2026,
author = { Daoud, Bilel AND Ding, Shanli AND Liu, Yi AND Khalfallah, Ali},
title = { { Structure-State Conditioned FB-to-DIBH Transition Learning for Breast Cancer Radiotherapy Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}