Efficient feature extraction using light-weight CNN attention-based deep learning architectures for ultrasound fetal...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Obstetrics

Background:

• Fetal plane classification (FPC) in ultrasound is crucial but challenging for clinicians.
• Accurate FPC requires identifying subtle fetal anatomical features, often time-consuming.
• Existing methods may lack efficiency and ease of deployment in clinical settings.

Purpose of the Study:

• To develop a lightweight AI model for accurate and efficient fetal plane classification.
• To assist obstetricians in identifying key fetal anatomical planes from ultrasound images.
• To enable real-time FPC for improved prenatal diagnostics.

Main Methods:

• Utilized a lightweight convolutional neural network architecture with attention mechanisms.
• Fine-tuned EfficientNet backbones pre-trained on ImageNet1k for feature extraction.
• Employed 3-layer perceptrons for classification and GradCAM for interpretation.

Main Results:

• Achieved high performance with Top-1 accuracy of 96.25% and Top-2 accuracy of 99.80%.
• Obtained an F1-Score of 0.9576 for fetal plane classification.
• The model features 40x fewer trainable parameters than existing benchmarks.

Conclusions:

• The proposed AI model offers a highly accurate and efficient solution for fetal plane classification.
• Its lightweight design facilitates deployment on edge devices for real-time clinical assistance.
• The model aids in diagnostics and treatment planning for expectant mothers.