Fetal Skull Reconstruction via Deep Convolutional Autoencoders.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Obstetrics

Background:

• Ultrasound (US) imaging is a primary tool for fetal screening.
• Three-dimensional ultrasound (3DUS) offers improved diagnostic capabilities over 2DUS but faces challenges with fetal movement and operator limitations.
• Acquiring complete fetal anatomical data can be hindered by occlusions in 3DUS volumes.

Purpose of the Study:

• To develop and evaluate deep convolutional neural networks for reconstructing fetal skulls from partially occluded 3DUS data.
• To address limitations in 3DUS acquisition caused by fetal movement and operator-dependent inspection.
• To improve the completeness and accuracy of fetal skull imaging in obstetric sonography.

Main Methods:

• Two deep convolutional architectures were proposed: a transfer learning deep convolutional network (TL-Net) and a conditional variational autoencoder (CVAE).
• The networks were trained and tested on 3DUS volumes with varying degrees of fetal skull occlusion, up to 50%.
• Performance was quantitatively assessed using metrics such as the Dice coefficient.

Main Results:

• Both TL-Net and CVAE demonstrated accurate fetal skull reconstruction even with significant occlusion (up to 50%).
• High accuracy was achieved when only 60% of the skull was visible, with Dice coefficients of 0.84±0.04 for CVAE and 0.83±0.03 for TL-Net.
• The proposed networks effectively handled partially occluded fetal skull data.

Conclusions:

• The developed deep learning models can reconstruct fetal skulls from incomplete 3DUS volumes.
• These reconstruction networks have the potential to optimize ultrasound acquisition protocols in obstetrics.
• The methods can reduce scan times and provide comprehensive anatomical information from challenging, partially occluded images.