Bridging the Gap: Semi-Automated 3D Histopathology-to-Ultrasound Registration

Towards Intraoperative Margin Assessment in Oral Cancer Surgery

Master Thesis (2026)
Author(s)

J.M.J. van Beurden (TU Delft - Mechanical Engineering)

Contributor(s)

Theo van Walsum – Mentor (Erasmus MC)

Maarten van Alphen – Mentor (Nederlands Kanker Instituut - Antoni van Leeuwenhoek ziekenhuis)

Luc Karssemakers – Mentor (Nederlands Kanker Instituut - Antoni van Leeuwenhoek ziekenhuis)

Andrea Borghesi – Mentor (Nederlands Kanker Instituut - Antoni van Leeuwenhoek ziekenhuis)

Lotje Zuur – Graduation committee member (Nederlands Kanker Instituut - Antoni van Leeuwenhoek ziekenhuis)

Faculty
Mechanical Engineering
More Info
expand_more
Publication Year
2026
Language
English
Graduation Date
27-08-2026
Awarding Institution
Delft University of Technology
Programme
Technical Medicine
Faculty
Mechanical Engineering
Downloads counter
5
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

Introduction: Surgical resection is the primary treatment for tongue squamous cell carcinoma (TSCC), but inadequate margins occur in up to 85\% of cases in Dutch cohorts. Histopathology is the gold standard for margin assessment, but its result only becomes available days after surgery, leaving the surgeon without intra-operative feedback. A deep learning model segmenting the tumour in ex-vivo 3D ultrasound (US) could give intra-operative feedback, but needs training labels with a known tumour boundary in US space. Histopathology provides that boundary in a separate coordinate system, so transferring it requires image registration.

Methods: A semi-automated 3D-to-3D registration pipeline was developed for oral-cavity specimens. A 3D histopathology volume is reconstructed from serial 2D H\&E slices, refined per slice using internal fiducial needles, and registered as a whole to the 3D US volume, held as the fixed reference. Registration estimates similarity transforms by aligning the specimen boundaries through signed distance fields. The pipeline was validated on a development and a held-out cohort without parameter re-tuning, using fiducial TRE, landmark TRE, NSD@1mm, and Dice; individual components were assessed through ablation and sensitivity experiments.

Results: On the development cohort ($n=5$), the pipeline reached a mean fiducial TRE of 1.56~mm, a mean landmark TRE of 1.69~mm, an NSD@1mm of 0.89, and a whole-specimen Dice of 0.80. On the held-out cohort ($n=7$), accuracy remained in similar range (2.13~mm, 2.23~mm, 0.87, 0.74).

Conclusion: To our knowledge, this is the first pipeline to register a reconstructed 3D histopathology volume to 3D US for oral-cavity specimens, with slice correspondence recovered semi-automatically rather than fixed by hand for each specimen. The current accuracy of roughly 2~mm limits direct margin assessment at the 1~mm threshold, but provides a scalable framework and an accuracy baseline for generating registered training data.

Files

License info not available