Calibration-Aware Semi-Supervised Fetal Head Segmentation with Boundary-Positive Contrast

Ufaq Khan, Umair Nawaz, Tajamul Ashraf, Tausifa Jan Saleem, Massimo Caputo, Srinivas Ananth Narayan, Muhammad Bilal, Junaid Qadir, Muhammad Haris Khan

Keywords: Fetal Segmentation, Semi-Supervised Learning, Boundary-Aware Learning

TL;DR: UltraSemiNet is a semi-supervised fetal head ultrasound segmentation method that selectively trusts calibrated predictions to get more accurate head boundaries.

Abstract: Accurate fetal head segmentation in ultrasound is hard to scale as labels are scarce and most errors occur at the head--background interface under speckle, shadowing, and low contrast. We present UltraSemiNet, a teacher--student framework that makes cross--pseudo supervision (CPS) selective via temperature calibration and a dual gate requiring high confidence and test-time augmentation (TTA) stability. We also introduce two boundary-focused modules that complement CPS: SAT, a boundary-positive spatial contrast that learns through ambiguous edges using an entropy belt and a soft-IoU agreement test; and \textbf{PCM}, a prototype-guided curriculum that maintains uncertainty-weighted head/background prototypes and targets feature-prototype discrepancies. Across two datasets (FBUI and HC18), UltraSemiNet improves overlap and boundary metrics over a calibrated CPS baseline (e.g., Dice $0.927{\rightarrow}0.971$; HD95 $7.9{\rightarrow}6.8$\,px), with similar cross-dataset trends. Crucially, the calibrated gate reduces miscalibration of the accepted pseudo-labels: both expected calibration error (ECE) and Brier score decrease overall, with the largest gains within the 0--2\,px boundary band, alongside improvements in pseudo-label accuracy. Ablations show CPS calibration, SAT, and PCM are complementary and concentrate improvements on boundary-sensitive metrics. In a blinded study, UltraSemiNet achieved better segmentation performance than two senior fetal medicine experts when evaluated against the dataset reference masks, indicating the potential to reduce manual refinements. Our code can be accessed on \href{https://github.com/Ufaqkhan/UltraSemiNet}{UltraSemiNet}

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Other

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 235