Sparse Convolutional Neural Network for High-Resolution Skull Shape Completion and Shape Super-Resolution

Our paper describes a practical solution to the curse of dimensionality in medical image analysis. The proposed approach is particularly relevant if your GPU memory does not possess the capacity to process the medical images at their original resolution and/or sluggish training prohibits efficient hyper-parameter tuning. Our work describes the utility of sparse convolutions in shape completion, super-resolution, and segmentation tasks. Experiments show that the proposed method can process high-resolution medical images using moderate memory and at a high speed.

---

💀 Skull Shape Completion and Super-Resolution

Thanks to sparse convolutions, a deep neural network can be trained on full-resolution skull images (512x512xZ) for shape completion and shape super-resolution tasks.

Previous approaches (or this) use dense convolutions, meaning images have to be downsampled to fit into GPU memory. A super-resolution network upsamples a coarse image to a higher resolution (e.g., 512x512xZ) and restores its fine geometric details on the shape surface.

Shape Completion (Input - Prediction - GT)	Super-Resolution (64 - 128 - 256 - 512)

---

🩻 Medical Image Segmentation

A detailed workflow of using sparse neural nets in medical image segmentation can be found here.

Segmentation 1	Segmentation 2

---

📝 Citation

If you find this work useful, please cite:

@article{li2023sparse,
  title={Sparse Convolutional Neural Network for High-resolution Skull Shape Completion and Shape Super-resolution},
  author={Li, Jianning and Gsaxner, Christina and Pepe, Antonio and Schmalstieg, Dieter and Kleesiek, Jens and Egger, Jan},
  journal={Scientific Reports},
  volume={13},
  doi={https://doi.org/10.1038/s41598-023-47437-6},
  year={2023}
}