GeezerVison

A few sentences describing the changes proposed in this pull request.

**Ready/Work in progress/Hold**

- [ ] Notebook runs automatically `./runner [-p <regex_pattern>]`

},

"source": [

"## Check transforms in DataLoader\n",

"Visulize a single batch to check the transforms."

"Visualize a single batch to check the transforms."

]

},

{

"## Quickly load data with DecathlonDataset\n",

"\n",

"Here we use `DecathlonDataset` to automatically download and extract the dataset.\n",

"It inherits MONAI `CacheDataset`, if you want to use less memory, you can set `cache_num=N` to cache N items for training and use the defaut args to cache all the items for validation, it depends on your memory size."

"It inherits MONAI `CacheDataset`, if you want to use less memory, you can set `cache_num=N` to cache N items for training and use the default args to cache all the items for validation, it depends on your memory size."

]

},

{

"# 3D Brain Tumor Segmentation with Swin UNETR (BraTS 21 Challenge)\n",

"\n",

"\n",

"This tutorial uses the [Swin UNETR](https://arxiv.org/pdf/2201.01266.pdf/) [1,2] model for the task of brain tumor segmentation using the [BraTS 21](http://braintumorsegmentation.org/) challenge dataset [3,4,5,6]. Swin UNETR ranked among top-perfoming models in BraTS 21 validation phase. The architecture of Swin UNETR is demonstrated as below\n",

"This tutorial uses the [Swin UNETR](https://arxiv.org/pdf/2201.01266.pdf/) [1,2] model for the task of brain tumor segmentation using the [BraTS 21](http://braintumorsegmentation.org/) challenge dataset [3,4,5,6]. Swin UNETR ranked among top-performing models in BraTS 21 validation phase. The architecture of Swin UNETR is demonstrated as below\n",

"\n",

"\n",

"The following features are included in this tutorial:\n",

"\n",

"\n",

"\n",

"## Refrences\n",

"## References\n",

"\n",

"\n",

"If you find this tutorial helpful, please consider citing [1] and [2]:\n",

"[2]: Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V. and Hatamizadeh, A., 2022. Self-supervised pre-training of swin transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 20730-20740).\n",

"\n",

"\n",

"### BraTS Dataset Refrences\n",

"### BraTS Dataset References\n",

"\n",

"[3] U.Baid, et al., The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification, arXiv:2107.02314, 2021.\n",

"\n",

"1. Construct segmentation and registration models.\n",

"2. Use the small number of segmentations (perhaps just one!) to pre-train the segmentation model. It's okay if it's not great.\n",

"3. Use the segmentation model as a source of weak supervision to help train the registration model.\n",

"4. Use the registration model as a soruce of weak supervision to help train the segmentation model.\n",

"4. Use the registration model as a source of weak supervision to help train the segmentation model.\n",

"5. Repeat steps 3 and 4 in alternation, or carry out steps 3 and 4 simultaneously in one joint training loop.\n",

"\n",

"The DeepAtlas method is fairly generic and does not need to prescribe specific model architectures\n",

"\n",

"---\n",

"\n",

"Note that we already have a segmentaton for each image. In the case of OASIS-1 data, these segmentations were originally done by [an algorithm](https://pubmed.ncbi.nlm.nih.gov/11293691/).\n",

"Note that we already have a segmentation for each image. In the case of OASIS-1 data, these segmentations were originally done by [an algorithm](https://pubmed.ncbi.nlm.nih.gov/11293691/).\n",

"In the case of synthetic data, true segmentations are generated for free by the process that creates the images.\n",

"\n",

"The situation in which DeepAtlas is interesting is one where we have many unlabled images and just a handful of labels. We will simulate this situation by pretending that only a few of the segmentation labels exist."

")\n",

"\n",

"\n",

"# Supress the many warnings related to depracation of the Analyze file format\n",

"# Supress the many warnings related to deprecation of the Analyze file format\n",

"# (without this, we would see warnings when the LoadImage transform calls itk to load Analyze files)\n",

"itk.ProcessObject.SetGlobalWarningDisplay(False)\n",

"\n",

"<a id=\"warp_warning\"></a>\n",

"We will need to be careful about when to use `warp` and when to use `warp_nearest`, especially with segmentation labels. Potential pitfalls:\n",

"- Using _trilinear interpolation_ to warp a segmentation with one channel containing class labels is a disaster: interpolating among labels would lead to intermediate values that aren't labels!\n",

"- Using _nearest neighbor inteprolation_ to warp a segmentation results in a reasonable output, but it is not a differentiable operation with respect to the deformation vector field (or, rather, derivatives vanish almost everywhere). This can be problematic for training, depending on what one is trying to do.\n",

"- Using _nearest neighbor interpolation_ to warp a segmentation results in a reasonable output, but it is not a differentiable operation with respect to the deformation vector field (or, rather, derivatives vanish almost everywhere). This can be problematic for training, depending on what one is trying to do.\n",

"- Warping a segmentation with `padding_mode=\"zeros\"` can pull in zero vectors, which might not make sense depending on how the segmentation is encoded. This is why we chose `padding_mode=\"border\"`."

]

},

bundle_inference_config.update({Const.KEY_INFERENCE_DATASET_DATA: test_datalist})

# update postprocessing with mean ensemble or vote ensemble

post_tranform = bundle_inference_config.config['postprocessing']

ensemble_tranform = {

post_transform = bundle_inference_config.config['postprocessing']

ensemble_transform = {

"_target_": f"{ensemble}Ensembled",

"keys": ["pred", "pred", "pred", "pred", "pred"],

"output_key": "pred"

}

if ensemble == 'Mean':

post_tranform["transforms"].insert(0, ensemble_tranform)

post_transform["transforms"].insert(0, ensemble_transform)

elif ensemble == 'Vote':

post_tranform["transforms"].insert(-1, ensemble_tranform)

post_transform["transforms"].insert(-1, ensemble_transform)

else:

raise NotImplementedError

bundle_inference_config.update({Const.KEY_INFERENCE_POSTPROCESSING: post_tranform})

bundle_inference_config.update({Const.KEY_INFERENCE_POSTPROCESSING: post_transform})

# update network weights

_networks = [bundle_inference_config.get_parsed_content("network")] * 5

Load the pre-trained weights from a MONAI model-zoo model and execute the transfer-learning in a typical PyTroch training program based on it.

Load the pre-trained weights from a MONAI model-zoo model and execute the transfer-learning in a typical PyTorch training program based on it.

This tutorial shows how to load a pretrained U-net from MONAI model-zoo and train it on BTCV dataset using pytorch. The pretrained U-net was trained on volumetric (3D) segmentation of CT spleen from MSD dataset.

"id": "f2596d86",

"metadata": {},

"source": [

"Note that with `axial_inferer` and `coronal_inferer`, the number of inference iterations is 64 and 256 repectively."

"Note that with `axial_inferer` and `coronal_inferer`, the number of inference iterations is 64 and 256 respectively."

]

}

],

" [\n",

" SqueezeDimd(keys=[\"img\", \"seg\"], dim=-1), # squeeze the last dim\n",

" Resized(keys=[\"img\", \"seg\"], spatial_size=[48, 48]),\n",

" # to use crop/pad instead of reszie:\n",

" # to use crop/pad instead of resize:\n",

" # ResizeWithPadOrCropd(keys=[\"img\", \"seg\"], spatial_size=[48, 48], mode=\"replicate\"),\n",

" ]\n",

")\n",