Using deep-learning algorithms to classify fetal brain ultrasound images as normal or abnormal.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Fetal Medicine

Background:

• Fetal brain abnormalities require accurate and timely diagnosis.
• Standard ultrasound interpretation can be challenging and subjective.
• Advancements in deep learning offer potential for automated image analysis.

Purpose of the Study:

• To assess the feasibility of deep learning for classifying fetal brain ultrasound images.
• To evaluate the performance of deep learning algorithms in segmenting and classifying fetal brain images.
• To explore the utility of deep learning in localizing lesions in abnormal fetal brain scans.

Main Methods:

• Utilized a large hospital database of 10,251 normal and 2,529 abnormal fetal brain ultrasound images.
• Trained deep learning algorithms for image segmentation, normal/abnormal classification, and lesion localization.
• Tested algorithm performance using precision, recall, Dice's coefficient, sensitivity, specificity, and expert evaluation of heat maps.

Main Results:

• Achieved high segmentation performance: precision (97.9%), recall (90.9%), Dice's coefficient (94.1%).
• Demonstrated excellent classification accuracy (96.3%) with high sensitivity (96.9%) and specificity (95.9%).
• Successfully localized lesions in 61.6% of abnormal images precisely, with an area under the ROC curve of 0.989.

Conclusions:

• Deep learning algorithms are feasible for segmenting and classifying fetal brain ultrasound images.
• These algorithms can effectively localize lesions, aiding in the diagnosis of fetal intracranial abnormalities.
• The study provides a foundation for future research in differential diagnosis of fetal brain conditions.