Sparse convolutional neural network for high-resolution skull shape completion and shape super-resolution.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Traditional convolutional neural networks (CNNs) struggle with spatially sparse data due to reliance on dense tensors.
• High-resolution 3D shapes, like medical scans, are often represented by sparse voxel grids, leading to computational inefficiency.
• Processing sparse medical data requires methods that minimize memory and computation overhead.

Purpose of the Study:

• To propose and evaluate a novel CNN model utilizing sparse tensors for efficient processing of high-resolution sparse data.
• To apply the sparse CNN to clinically relevant skull reconstruction tasks: shape completion and super-resolution.
• To demonstrate the method's superiority over dense CNN counterparts in terms of performance and resource efficiency.

Main Methods:

• Developed a CNN architecture that operates on sparse tensors, processing only non-empty voxels.
• Evaluated the sparse CNN on 3D skull data for reconstruction tasks, comparing it against dense CNN approaches.
• Analyzed memory consumption during training and inference concerning image resolution and voxel count.

Main Results:

• The sparse CNN significantly outperformed dense CNNs in quantitative and qualitative skull reconstruction.
• Substantially reduced memory requirements for both training and inference compared to dense CNNs.
• Memory consumption of the sparse CNN demonstrated near-linear growth with resolution during inference and minimal increase during training.

Conclusions:

• Sparse CNNs are highly effective for processing spatially sparse medical data, particularly in 3D shape reconstruction.
• The proposed method offers a more efficient alternative to dense CNNs for high-resolution medical imaging tasks.
• The findings are applicable to other spatially sparse problems, as evidenced by successful tests on aorta and heart datasets.