Contrast and Gain-Aware Attention: A Plug-and-Play Feature Fusion Attention Module for Torso Region Fetal Plane Identification

Abstract

Accurate identification of fetal torso ultrasound planes is essential in pre-natal examinations, as it plays a critical role in the early detection of severe fetal malformations and this process is heavily dependent on the clinical expertise of health care providers. However, the limited number of medical professionals skilled at identification and the complexity of fetal plane screening underscore the need for efficient diagnostic support tools. Clinicians often encounter challenges such as image artifacts and the intricate nature of fetal planes, which require adjustments to image gain and contrast to obtain clearer diagnostic information. In response to these challenges, we propose the contrast and gain-aware attention mechanism. This method generates images under varying gain and contrast conditions, and utilizes an attention mechanism to mimic the clinician’s decision-making process. The system dynamically allocates attention to images based on these conditions, integrating feature fusion through a lightweight attention module. Positioned in the first layer of the model, this module operates directly on images with different gain and contrast settings. Here we integrated this attention mechanism into ResNet18 and ResNet34 models to predict key fetal torso planes: the transverse view of the abdomen, the sagittal view of the spine, the transverse view of the kidney and the sagittal view of the kidney. Our experimental results showed that this approach significantly enhances performance compared with traditional models, with minimal addition to model parameters, ensuring both efficiency and effectiveness in fetal torso ultrasound plane identification. Our codes are available at https://github.com/sysll/CCGAA.