Fetal Cerebellum Landmark Detection Based on 3D MRI: Method and Benchmark

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Fetal cerebellum landmark detection is vital for evaluating fetal brain development.
• Current deep learning methods often rely on 2D ultrasound or thick 3D MRI, limiting accuracy.
• 3D fetal brain images present challenges like noise and fuzzy boundaries, hindering traditional detection methods.

Purpose of the Study:

• To develop an accurate method for fetal cerebellum landmark detection on thin 3D MRIs.
• To introduce a novel deep learning network, the Anatomical Pseudo-label Guided Attention (APGA) network.
• To establish a 3D MRI-based benchmark dataset for fetal cerebellum landmark detection.

Main Methods:

• The APGA network utilizes a shared encoder with two decoders for landmark regression and anatomical pseudo-label segmentation.
• A Feature Decoupling Transformer (FDT) is integrated into the encoder to enhance feature calibration for both tasks.
• The method requires only the encoder, FDT, and landmark decoder during inference.

Main Results:

• Extensive experiments demonstrated the effectiveness of the APGA network on a proposed benchmark and an out-of-domain test set.
• The study confirmed that 3D biometrics provide superior results compared to 2D biometrics for fetal cerebellum analysis.
• The APGA network achieved improved accuracy in fetal cerebellum landmark detection using 3D MRI data.

Conclusions:

• The APGA network offers a robust solution for accurate fetal cerebellum landmark detection on 3D MRIs.
• The developed 3D MRI-based benchmark facilitates further research in this area.
• The findings highlight the superiority of 3D imaging and advanced deep learning techniques for fetal neurodevelopment assessment.