3365 update highlight 0.8

docs/source/highlights.md

@@ -163,20 +163,31 @@ If the pipeline includes random transformations, users may want to observe the e
 [Invert transforms and TTA tutorials](https://github.com/Project-MONAI/tutorials/blob/master/modules/inverse_transforms_and_test_time_augmentations.ipynb) introduce details about the API with usage examples.
 
 (1) The last column is the inverted data of model output:
+
 ![invert transform](../images/invert_transforms.png)
 
 (2) The TTA results of `mode`, `mean` and `standard deviation`:
+
 ![test time augmentation](../images/tta.png)
 
+### 15. Visualization of transform examples
+To help clearly introduce the transform functionalities, MONAI provides visualization examples in the [API document](https://docs.monai.io/en/latest/transforms.html) for almost every transform, including spatial transforms, intensity transforms, crop / pad transforms, etc.
+
+For example:
+
+![rand gaussian noise](../images/rand_gaussian_noise.png)
+
 ## Datasets and DataLoader
 ### 1. Cache IO and transforms data to accelerate training
 Users often need to train the model with many (potentially thousands of) epochs over the data to achieve the desired model quality. A native PyTorch implementation may repeatedly load data and run the same preprocessing steps for every epoch during training, which can be time-consuming and unnecessary, especially when the medical image volumes are large.
 
 MONAI provides a multi-thread `CacheDataset` and `LMDBDataset` to accelerate these transformation steps during training by storing the intermediate outcomes before the first randomized transform in the transform chain. Enabling this feature could potentially give 10x training speedups in the [Datasets experiment](https://github.com/Project-MONAI/tutorials/blob/master/acceleration/dataset_type_performance.ipynb).
+
 ![digital pathology](../images/cache_dataset.png)
 
 ### 2. Cache intermediate outcomes into persistent storage
 The `PersistentDataset` is similar to the CacheDataset, where the intermediate cache values are persisted to disk storage or LMDB for rapid retrieval between experimental runs (as is the case when tuning hyperparameters), or when the entire data set size exceeds available memory. The `PersistentDataset` could achieve similar performance when comparing to `CacheDataset` in [Datasets experiment](https://github.com/Project-MONAI/tutorials/blob/master/acceleration/dataset_type_performance.ipynb).
+
 ![cachedataset speed](../images/datasets_speed.png)
 
 ### 3. SmartCache mechanism for big datasets
@@ -227,6 +238,7 @@ To quickly get started with popular training data in the medical domain, MONAI p
 MONAI always welcome new contributions of public datasets, please refer to existing Datasets and leverage the download and extracting APIs, etc. [Public datasets tutorial](https://github.com/Project-MONAI/tutorials/blob/master/modules/public_datasets.ipynb) indicates how to quickly set up training workflows with `MedNISTDataset` and `DecathlonDataset` and how to create a new `Dataset` for public data.
 
 The common workflow of predefined datasets:
+
 ![pre-defined dataset](../images/dataset_progress.png)
 
 ### 7. Partition dataset for cross validation
@@ -238,7 +250,9 @@ In addition to the regular preprocessing transform while loading, it also suppor
 
 ### 9. `ThreadDataLoader` vs. `DataLoader`
 If the transforms are light-weighted, especially when we cache all the data in RAM, the multiprocessing of PyTorch `DataLoader` may cause unnecessary IPC time and cause the drop of GPU utilization after every epoch. MONAI provides `ThreadDataLoader` which executes the transforms in a separate thread:
+
 ![threaddataloader](../images/threaddataloader.png)
+
 a `ThreadDataLoader` example is available at [Spleen fast training tutorial](https://github.com/Project-MONAI/tutorials/blob/master/acceleration/fast_training_tutorial.ipynb).
 
 ## Losses
@@ -248,6 +262,7 @@ There are domain-specific loss functions in the medical imaging research which a
 MONAI provides several advanced features in optimizers to help accelerate the training or fine-tuning progress. For example, `Novograd` optimizer can be used to converge faster than the traditional optimizers. And users can easily define different learning rates for the model layers based [on the `generate_param_groups` utility API](https://github.com/Project-MONAI/tutorials/blob/master/modules/layer_wise_learning_rate.ipynb).
 
 Another important feature is `LearningRateFinder`. The learning rate range test increases the learning rate in a pre-training run between two boundaries in a linear or exponential manner. It provides valuable information on how well the network can be trained over a range of learning rates and what the optimal learning rates are. [LearningRateFinder tutorial](https://github.com/Project-MONAI/tutorials/blob/master/modules/learning_rate.ipynb) indicates the API usage examples.
+
 ![learning rate finder plot](../images/lr_finder.png)
 
 ## Network architectures
@@ -269,7 +284,7 @@ add_module('conv1', conv_type(in_channels, out_channels, kernel_size=1, bias=Fal
 ```
 
 ### 2. Implementation of generic 2D/3D networks
-And there are several 1D/2D/3D-compatible implementations of intermediate blocks and generic networks, such as UNet, DynUNet, DenseNet, GAN, AHNet, VNet, SENet(and SEResNet, SEResNeXt), SegResNet, EfficientNet, Attention-based transformer networks. All the networks can support PyTorch serialization pipeline based on `torch.jit.script`.
+And there are several 1D/2D/3D-compatible implementations of intermediate blocks and generic networks, such as UNet, DynUNet, DenseNet, GAN, AHNet, VNet, SENet(and SEResNet, SEResNeXt), SegResNet, EfficientNet, Attention-based transformer networks, Multi-instance learning networks, DiNTS for AutoML, etc. All the networks can support PyTorch serialization pipeline based on `torch.jit.script`.
 
 ### 3. Network adapter to finetune final layers
 Instead of training from scratch, we often leverage the existing models, and finetune the final layers of a network for new learning tasks. MONAI provides a `NetAdapter` to easily replace the last layer of a model by a convolutional layer or a fully-connected layer. A typical usage example is to adapt [Torchvision models trained with ImageNet](https://pytorch.org/vision/stable/models.html) for other learning tasks.
@@ -302,10 +317,22 @@ With a `Cumulative` base class, intermediate metric outcomes can be automaticall
 
 ### 3. Metrics report generation
 During evaluation, users usually save the metrics of every input image, then analyze the bad cases to improve the deep learning pipeline. To save detailed information of metrics, MONAI provided a handler `MetricsSaver`, which can save the final metric values, raw metric of every model output channel of every input image, metrics summary report of operations: `mean`, `median`, `max`, `min`, `<int>percentile`, `std`, etc. The `MeanDice` reports of validation with prostate dataset are as below:
+
 ![metrics report example](../images/metrics_report.png)
 
 ## Visualization
-Beyond the simple point and curve plotting, MONAI provides intuitive interfaces to visualize multidimensional data as GIF animations in TensorBoard. This could provide a quick qualitative assessment of the model by visualizing, for example, the volumetric inputs, segmentation maps, and intermediate feature maps. A runnable example with visualization is available at [UNet training example](https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/torch/unet_training_dict.py).
+Beyond the simple point and curve plotting, MONAI provides intuitive interfaces to visualize multidimensional data as GIF animations in TensorBoard. This could provide a quick qualitative assessment of the model by visualizing, for example, the volumetric inputs, segmentation maps, and intermediate feature maps. A runnable example with visualization is available at [UNet training example](https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/torch/unet_training_dict.py). To work with ignite program, MONAI also provides several ignite handlers to visualize training curve and metrics with `TensorBoard` or `MLFlow`, more details is available in [TensorBoard and MLFlow handlers example](https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/unet_segmentation_3d_ignite.ipynb).
+
+To easily visualize a 3D image as frames of 2D images, MONAI provides the utility `matshow3d` based on `matplotlib` library. It can plot frames of image for the specified dimension, showing a spleen 3D image as example:
+`matshow3d(volume=image, figsize=(100, 100), every_n=10, frame_dim=-1 show=True, cmap="gray")`
+
+![matshow3d example](../images/matshow3d.png)
+
+MONAI also provides the `blend_images` utility to blend the `image` and `label` to a RGB color image to better visualize the segmentation regions with the specified `cmap` mode and weights, etc. Showing a spleen segmentation `image` and the corresponding `label` as example:
+
+![blend example](../images/blend.png)
+
+For more details of `TensorBoard utility`, `matshow3d` and `blend_images`, please check the [visualziation tutorial](https://github.com/Project-MONAI/tutorials/blob/master/modules/transform_visualization.ipynb).
 
 And to visualize the class activation mapping for a trained classification model, MONAI provides CAM, GradCAM, GradCAM++ APIs for both 2D and 3D models:
 
@@ -384,10 +411,26 @@ G. Wang, X. Liu, C. Li, Z. Xu, J. Ruan, H. Zhu, T. Meng, K. Li, N. Huang, S. Zha
 [A reimplementation](https://monai.io/research/lamp-automated-model-parallelism) of the LAMP system originally proposed by:
 
 Wentao Zhu, Can Zhao, Wenqi Li, Holger Roth, Ziyue Xu, and Daguang Xu (2020) "LAMP: Large Deep Nets with Automated Model Parallelism for Image Segmentation." MICCAI 2020 (Early Accept, paper link: https://arxiv.org/abs/2006.12575)
+
 ![LAMP UNet](../images/unet-pipe.png)
 
+### 3. DiNTS: Differentiable Neural Network Topology Search for 3D Medical Image Segmentation
+MONAI integrated the `DiNTS` module to support more flexible topologies and joint two-level search. It provides a topology guaranteed discretization algorithm and a discretization aware topology loss for the search stage to minimize the discretization gap, and a cost usage aware search method which can search 3D networks with different GPU memory requirements. For more details, please check the [DiNTS tutorial](https://monai.io/research/dints.html).
+
+![DiNTS](../images/dints-overview.png)
+
+### 4. Accounting for Dependencies in Deep Learning Based Multiple Instance Learning for Whole Slide Imaging
+For [classification of digital pathology whole slide images (WSI)](https://arxiv.org/abs/2111.01556), MONAI introduces new transforms and network modules for multiple instance learning. These include self-attention transformer blocks for explicitly accounting of the dependencies between instances (image patches) during training. For more details, please check out the [multiple instance learning tutorial](https://github.com/Project-MONAI/tutorials/tree/master/pathology/multiple_instance_learning). ![multi-instance](../images/mil-patches.jpg)
+
+### 5. Self-supervised representation learning
+MONAI starts to explore self-supervised representation learning in this milestone release. The Vision Transformer has been extended to learn from self-supervised reconstruction tasks with various data augmentation and a regularized contrastive loss. The weights of the pre-trained backbone could be used to enhance the performance of the novel downstream deep learning tasks.
+
+The [tutorial](https://github.com/Project-MONAI/tutorials/tree/master/self_supervised_pretraining) shows how to generate a good set of pre-trained weights using unlabeled data with self-supervised tasks, then use the pre-trained weights to perform fine-tuning on a fully supervised volumetric segmentation task using a transformer based `UNETR`.
+
+![self-supervised](../images/ssl_overview.png)
+
 ## Performance optimization and GPU acceleration
-Typically, model training is a time-consuming step during deep learning development, especially in medical imaging applications. Volumetric medical images are usually large (as multi-dimensional arrays) and the model training process can be complex. Even with powerful hardware (e.g. CPU/GPU with large RAM), it is not easy to fully leverage them to achieve high performance. MONAI provides a fast training guide to achieve the best performance: https://github.com/Project-MONAI/tutorials/blob/master/acceleration/fast_model_training_guide.md.
+Typically, model training is a time-consuming step during deep learning development, especially in medical imaging applications. Volumetric medical images are usually large (as multi-dimensional arrays) and the model training process can be complex. Even with powerful hardware (e.g. CPU/GPU with large RAM), it is not easy to fully leverage them to achieve high performance. MONAI provides [a fast training guide to achieve the best performance](https://github.com/Project-MONAI/tutorials/blob/master/acceleration/fast_model_training_guide.md).
 
 NVIDIA GPUs have been widely applied in many areas of deep learning training and evaluation, and the CUDA parallel computation shows obvious acceleration when comparing to traditional computation methods. To fully leverage GPU features, many popular mechanisms raised, like automatic mixed precision (AMP), distributed data parallel, etc. MONAI can support these features and provides rich examples.
 

docs/source/whatsnew_0_8.md

@@ -14,8 +14,8 @@ high search efficiency and budgeted memory usage.
 It provides a topology guaranteed discretization algorithm and a
 discretization-aware topology loss for the search stage to minimize the
 discretization gap. The module is memory usage aware and is able to search 3D
-networks with different GPU memory requirements. For more details, please checkout the
-[DiNTS tutorial](https://github.com/Project-MONAI/tutorials/tree/master/automl).
+networks with different GPU memory requirements. For more details, please check out the
+[DiNTS tutorial](https://monai.io/research/dints.html).
 
 ![DiNTS](../images/dints-overview.png)
 
@@ -25,7 +25,7 @@ For [classification of digital pathology whole slide images
 network modules for multiple instance learning. These include self-attention
 transformer blocks for explicitly accounting of the dependencies between instances
 (image patches) during training. For more details,
-please checkout the [multi-instance tutorial](https://github.com/Project-MONAI/tutorials/tree/master/pathology/multiple_instance_learning)
+please check out the [multiple instance learning tutorial](https://github.com/Project-MONAI/tutorials/tree/master/pathology/multiple_instance_learning).
 
 ![multi-instance](../images/mil-patches.jpg)
 
@@ -41,7 +41,7 @@ shows how to generate a good set of pre-trained weights using unlabeled data
 with self-supervised tasks, then use the pre-trained weights to perform
 fine-tuning on a fully supervised volumetric segmentation task using a transformer based `UNETR`.
 
-![self-supervised](../images/SSL_Overview_Figure.png)
+![self-supervised](../images/ssl_overview.png)
 
 ## Major usability improvements in `monai.transforms`
 `monai.transforms` are now more flexible and easy to use in version 0.8.