[Paper Review] Attention-Gated Networks for Improving Ultrasound Scan Plane Detection

Jo Schlemper, Ozan Oktay+6 more|arXiv (Cornell University)|Apr 15, 2018

Domain Adaptation and Few-Shot LearningComputer Science87 citations

TL;DR

The paper introduces Attention-Gated Sononet (AG-Sononet), a soft self-attention mechanism integrated into a fetal ultrasound scan plane detector to leverage local features, improve precision, and provide weakly supervised localisation with minimal overhead.

ABSTRACT

In this work, we apply an attention-gated network to real-time automated scan plane detection for fetal ultrasound screening. Scan plane detection in fetal ultrasound is a challenging problem due the poor image quality resulting in low interpretability for both clinicians and automated algorithms. To solve this, we propose incorporating self-gated soft-attention mechanisms. A soft-attention mechanism generates a gating signal that is end-to-end trainable, which allows the network to contextualise local information useful for prediction. The proposed attention mechanism is generic and it can be easily incorporated into any existing classification architectures, while only requiring a few additional parameters. We show that, when the base network has a high capacity, the incorporated attention mechanism can provide efficient object localisation while improving the overall performance. When the base network has a low capacity, the method greatly outperforms the baseline approach and significantly reduces false positives. Lastly, the generated attention maps allow us to understand the model's reasoning process, which can also be used for weakly supervised object localisation.

Motivation & Objective

• Improve real-time scan plane detection in fetal ultrasound despite low image quality and class imbalance.
• Incorporate self-gated soft-attention to exploit local information without bounding-box supervision.
• Enable visualisable attention maps for explainability and weakly supervised localisation.

Proposed method

• Extend Sononet with self-gated soft-attention blocks at selected layers to produce attention maps.
• Use a generalized attention formulation with linear transformations and gating to compute compatibility scores.
• Implement grid attention to preserve spatial context by using a gridded global feature representation and bilinear upsampling.
• Experiment with aggregation strategies across scales: independent per-scale predictions with averaging, deep supervision, and a fine-tuned joint predictor.
• Compare AG-Sononet variants against baseline Sononet across model capacities (8, 16, 32 initial filters).

Experimental results

Research questions

• RQ1 Does soft-attention improve accuracy and precision for fetal scan plane classification compared to a strong baseline?
• RQ2 Can attention maps provide weakly supervised localisation without bounding-box labels and with low computational overhead?
• RQ3 How do aggregation strategies across multi-scale features affect learning and performance?
• RQ4 What is the impact of model capacity on the benefit of attention gating in this task?

Key findings

• AG-Sononet consistently improves over Sononet across capacities, with higher precision and reduced false positives.
• Attention gating enables better discrimination of visually similar cardiac views (e.g., 4CH and 3VV) and improves kidney, profile, and spine classifications.
• Attention maps produced by AG-Sononet enable near real-time weakly supervised localisation without extra computational cost.
• Different aggregation strategies show that a stable, two-stage approach (per-scale training followed by joint fine-tuning) yields the best performance.
• AG-Sononet with 8/16/32 initial filters achieves competitive accuracy and F1 with fewer parameters than progressively larger baselines.