3D deep learning requires significantly higher computation power than sequential 2D image analyses. Image compression for 3D deep learning may compromise the results for lesion detection. Is there literature that you can recommend for the comparison of 2D and 3D deep learning as well as a dedicated opinion on what you prefer in regard to the detection and characterization of brain lesions?
Kind regards!
JB
The performance of 2D deep learning far exceeds that of 3D mainly because of the relatively small size of the available 3D datasets.
See this article for a review specific to MRI imaging: https://medium.com/stanford-ai-for-healthcare/dont-just-scan-this-deep-learning-techniques-for-mri-52610e9b7a85
Many works on 3D CNNs (e.g. OctNet, VoxNet) down-sample the 3D space into a grid of voxels of predefined size (voxelisation) to overcome the memory challenge when dealing with 3D data. However recent works (e.g. PointNet, PointNet++, SegCloud) have achieved point-level segmentation.
I think that 3D deep learning will be more effective than 2D in the future, due to the increased level of information in 3D data, if the efficiency of these algorithms continues to improve at the rate it is.
Yes. In 3D brain images, surely the depthness of the lesion can be measured accurately than in the 2D images using some geometrical methods.
Thank you for your opiniom! - For now, I conclude for myself that in the current clinical application, it appears to be a better standard to utilize 2D images instead of downsampling 3D images due to lack of computing power.
- Badges
- Science topic