[Paper Review] Weakly Supervised Estimation of Shadow Confidence Maps in Fetal Ultrasound Imaging

Qingjie Meng, Matthew Sinclair+12 more|arXiv (Cornell University)|Nov 20, 2018

Fetal and Pediatric Neurological Disorders25 references4 citations

TL;DR

This paper proposes a weakly supervised deep learning method to estimate dense, continuous shadow confidence maps in 2D fetal ultrasound images using only image-level labels and a small set of coarse pixel-wise annotations. By leveraging a shadow-seg module, transfer function, and confidence estimation network, the method achieves superior segmentation accuracy and more consistent confidence maps than human experts, enabling real-time guidance for sonographers and improved performance in downstream tasks like image classification and biometric measurement.

ABSTRACT

Detecting acoustic shadows in ultrasound images is important in many clinical and engineering applications. Real-time feedback of acoustic shadows can guide sonographers to a standardized diagnostic viewing plane with minimal artifacts and can provide additional information for other automatic image analysis algorithms. However, automatically detecting shadow regions using learning-based algorithms is challenging because pixel-wise ground truth annotation of acoustic shadows is subjective and time consuming. In this paper we propose a weakly supervised method for automatic confidence estimation of acoustic shadow regions. Our method is able to generate a dense shadow-focused confidence map. In our method, a shadow-seg module is built to learn general shadow features for shadow segmentation, based on global image-level annotations as well as a small number of coarse pixel-wise shadow annotations. A transfer function is introduced to extend the obtained binary shadow segmentation to a reference confidence map. Additionally, a confidence estimation network is proposed to learn the mapping between input images and the reference confidence maps. This network is able to predict shadow confidence maps directly from input images during inference. We use evaluation metrics such as DICE, inter-class correlation and etc. to verify the effectiveness of our method. Our method is more consistent than human annotation, and outperforms the state-of-the-art quantitatively in shadow segmentation and qualitatively in confidence estimation of shadow regions. We further demonstrate the applicability of our method by integrating shadow confidence maps into tasks such as ultrasound image classification, multi-view image fusion and automated biometric measurements.

Motivation & Objective

• To address the challenge of limited and subjective pixel-wise annotation for acoustic shadow detection in fetal ultrasound.
• To develop a weakly supervised method that generates dense, continuous confidence maps for shadow regions using only image-level labels and sparse coarse annotations.
• To improve the consistency and accuracy of shadow detection beyond human inter-observer variability.
• To enable real-time feedback for sonographers to optimize scanning planes and reduce artifacts.
• To enhance downstream applications such as image classification, multi-view fusion, and automated biometric measurements through integrated shadow confidence maps.

Proposed method

• A shadow-seg module is trained on images with global image-level labels ('has shadow' or 'shadow-free') to learn general shadow features.
• A transfer function maps the binary shadow segmentation output to a reference confidence map using mean intensity of true positive regions (Eq. 1).
• A confidence estimation network is trained to predict continuous shadow confidence maps directly from input images, using the reference maps as supervision.
• The method leverages a small set of coarse pixel-wise annotations (approx. 10% of image-level labeled images) to refine the latent space for fetal brain-specific shadow features.
• The network is trained end-to-end with loss functions including DICE, MSE, and soft DICE to optimize segmentation and confidence estimation.
• The approach is robust to noisy or inaccurate coarse annotations due to the use of intensity-based smoothing and evaluation metrics like ICC and soft DICE.

Experimental results

Research questions

• RQ1 Can a weakly supervised deep learning method generate accurate and consistent shadow confidence maps using only image-level labels and sparse coarse annotations?
• RQ2 How does the proposed method compare to human inter-observer variability in terms of shadow segmentation accuracy and confidence map consistency?
• RQ3 Can the estimated shadow confidence maps improve performance in downstream ultrasound applications such as image classification and biometric measurement?
• RQ4 Does the method generalize to detecting multiple disjoint shadow regions without relying on spatial correlation or heuristic hyperparameters?
• RQ5 Can the method provide real-time feedback for sonographers to guide optimal scanning plane selection?

Key findings

• The proposed method outperforms state-of-the-art methods in quantitative shadow segmentation, achieving higher DICE and lower MSE scores than existing approaches.
• The method demonstrates greater consistency than human annotations, with inter-class correlation (ICC) values indicating superior reliability across annotators.
• The confidence estimation network produces more visually plausible and continuous confidence maps than baseline methods, especially in regions with complex shadow patterns.
• The integration of shadow confidence maps improves performance in standard plane classification, multi-view image fusion, and automated biometric measurements, demonstrating utility in real-world applications.
• The method achieves significantly faster inference times than previous approaches, enabling real-time deployment for clinical guidance.
• The model generalizes well to detecting multiple disjoint shadow regions and shows robustness to imperfect coarse annotations due to intensity-based smoothing and loss design.