HDC: Hierarchical Distillation for Multi-level Noisy Consistency in Semi-Supervised Fetal Ultrasound Segmentation

Event: CVPR 2025 · Duration: 5 min · ▶ Watch on YouTube

Abstract

Accurate segmentation of fetal ultrasound images is crucial for prenatal diagnosis, but it is hindered by the high cost and expertise required for pixel-level annotations, as well as the inherent low contrast and shadow noise in ultrasound images. This work proposes HDC (Hierarchical Distillation for Multi-level Noisy Consistency), a single-teacher semi-supervised framework designed to address these challenges. HDC utilizes a dual-decoder student architecture with multi-level noise injection and hierarchical distillation losses, including Correlation Guidance Loss for feature alignment, Mutual Information Loss for decoder consistency, and Pixel-Level Consistency Loss for prediction refinement. The method aims to align feature distributions and achieve accurate segmentation with minimal labeled samples, demonstrating superior performance over existing methods on challenging fetal ultrasound datasets.

Speakers

  • Tran Quoc Khanh Le — University of Information Technology & AI VIETNAM

Talks (1)

  • 00:00:00 — Tran Quoc Khanh Le: HDC: Hierarchical Distillation for Multi-level Noisy Consistency in Semi-Supervised Fetal Ultrasound Segmentation
    • This presentation introduces HDC, a novel semi-supervised learning framework for fetal ultrasound segmentation that leverages hierarchical distillation and multi-level noise injection to improve robustness and accuracy with limited labeled data.

Key Takeaways

  • HDC proposes a single-teacher, dual-decoder student architecture for efficient semi-supervised fetal ultrasound segmentation.
  • The framework incorporates multi-level noise injection and hierarchical distillation losses, specifically Correlation Guidance Loss, Mutual Information Loss, and Pixel-Level Consistency Loss, to enhance model robustness and consistency.
  • Experimental results on the FUGC and PSFH datasets demonstrate that HDC achieves superior quantitative and qualitative performance compared to several state-of-the-art semi-supervised learning methods.
  • Ablation studies confirm the individual contributions of each proposed loss component to the overall improved performance of the HDC model.
  • HDC effectively addresses challenges like low contrast, shadow noise, and limited labeled data in medical imaging, making it a promising solution for real-world applications.

Methods / Models / Datasets Mentioned

  • HDC
  • Mean Teacher
  • FixMatch
  • CPS
  • PS-MT
  • CCVC
  • Dual Teacher
  • AD-MT
  • ResNet50
  • ResNet101
  • Correlation Guidance Loss (L_CG)
  • Mutual Information Loss (L_MI)
  • Pixel-Level Consistency Loss (Pix Loss)

Topics

Semi-supervised learning · Fetal ultrasound segmentation · Hierarchical distillation · Multi-level noise injection · Medical image analysis · Low contrast images · Shadow noise · Feature alignment · Decoder consistency · Pixel-level consistency

Notes

Open for commentary — connections to other work, critiques, follow-up reading.