Awesome
3DMedPT
<!-- **\[our code for IntrA and Modelnet40 classification is released]** 3D Medical Point Transformer: Introducing Convolution to Attention Networks for Medical Point Cloud Analysis [](https://paperswithcode.com/sota/classification-on-intra?p=3d-medical-point-transformer-introducing) [](https://paperswithcode.com/sota/3d-part-segmentation-on-intra?p=3d-medical-point-transformer-introducing) [[arxiv]](https://arxiv.org/pdf/2112.04863.pdf) [[project page]](https://3dmedpt.github.io/) Author: Jianhui Yu, Chaoyi Zhang, Heng Wang, Dingxin Zhang, Yang Song, Tiange Xiang, Dongnan Liu, Weidong Cai --> <!-- ## Model Architecture  -->Requirements
- Python 3.7
- Pytorch 1.6
- CUDA 10.0
- Packages: tqdm, sklearn, visualdl, einops, natsort
- To build the CUDA kernel for FPS:
<b>NOTE:</b> If you encounter problems while building the kernel, you can refer to Pointnet2_PyTorch for solutions.pip install pointnet2_ops_lib/.
Data
The IntrA dataset can be downloaded from intra3d2019, and you need to unzip the files to data/IntrA3D.
The ModelNet40 dataset is automatically downloaded.
Performance
- State-of-the-art accuracy on IntrA classification: <b>0.936</b> (F1 score)
- State-of-the-art accuracy on IntrA segmentation:
- IoU: <b>94.82%</b> on healthy vessel and <b>82.39%</b> on aneurysm
- DSC: <b>97.29%</b> on healthy vessel and <b>89.71%</b> on aneurysm
- ModelNet40 classification: <b>93.4%</b>
- ShapeNet40 part segmentation: <b>84.3%</b> (class mIoU)
Training Command
-
For IntrA model train (1024 points)
CUDA_VISIBLE_DEVICES=xx python main_intra.py --exp_name intra_cls_train --mode train --num_points 1024 --lr 0.01 --use_sgd True --num_K 32 64 -
For ModelNet40 model train (1024 points)
CUDA_VISIBLE_DEVICES=xx python main_m40.py --exp_name m40_cls_train --mode train --num_points 1024 --epochs 250 --lr 0.001 --num_K 20 20
<b>NOTE:</b> To achieve a fast computational speed, you can also uncomment torch.backends.cudnn.benchmark = True and comment out torch.backends.cudnn.deterministic = True, while the final results might vary.
Acknowledgement
Our code borrows from: