Decision Fusion-Based Fetal Ultrasound Image Plane Classification Using Convolutional Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Fetal Development

Background:

• Automated classification of fetal ultrasound images aids retrospective analysis and identification of developmental indicators.
• Current machine learning methods are limited to using either whole images or cropped patches, leading to misclassification due to similar organ appearances and lack of localized detail.

Purpose of the Study:

• To develop and validate a novel machine learning approach for accurate classification of 14 distinct fetal structures in 2-D ultrasound images.
• To improve upon existing methods by fusing global and local image information for enhanced classification and localization.

Main Methods:

• A novel method was developed that fuses information from both whole fetal ultrasound images and discriminant regions of fetal structures.
• Two feature extractors were trained using fine-tuned pre-trained convolutional neural networks on global images and localized regions.
• The approach integrates classification decisions from global and local features without relying on prior information and enables structure localization.

Main Results:

• The proposed method achieved a mean accuracy of 97.05%, mean precision of 76.47%, and mean recall of 75.41% on a dataset of 4074 images.
• A Cohen's kappa of 0.72 indicated substantial agreement between the method's predictions and ground truth.
• The method demonstrated statistically significant superiority (p < 0.05) over non-fusion-based approaches and achieved 92% accuracy on images without scanner overlays.

Conclusions:

• The developed fusion-based machine learning method accurately classifies multiple fetal structures in 2-D ultrasound images.
• This approach overcomes limitations of existing methods by integrating global and local image features, enhancing classification and enabling localization.
• The method shows significant potential for retrospective classification of ultrasound images in clinical research and improving fetal development assessment.