NVIDIA MONAI Workshop 1

This repository contains a series of hands-on tutorials for the NVIDIA MONAI Workshop, focusing on medical image analysis using MONAI and PyTorch.

1. MONAI Label with OHIF Viewer

Notebook: lab_1_MONAI_Label_with_OHIF_Viewer/lab_1_monailabel_radiology_OHIF.ipynb

This tutorial demonstrates the end-to-end setup of the web-based OHIF viewer and MONAI Label plugin for radiology applications.

Key Components:

• App: MONAI Label Radiology App
• Viewer: OHIF Viewer
• Server: Orthanc DICOM Server
• Task: Interactive Segmentation

Workflow:

1. Environment Setup: Install MONAI Label and dependencies.
2. App Download: Download the standard "radiology" app.
3. Data Management: Download Spleen CT DICOM data and upload it to the Orthanc DICOM server.
4. Server Launch: Start the MONAI Label server connected to Orthanc.
5. Interactive Segmentation: Use the OHIF viewer to perform auto-segmentation on the spleen using pre-trained models provided by MONAI Label.

2. 2D Diabetic Retinopathy Classification

Notebook: lab_2_Classification_2D/lab_2_2d_classification_dr.ipynb

This notebook demonstrates a 2D classification task to grade Diabetic Retinopathy (DR) from fundus images into five stages.

Key Components:

• Model: DenseNet121
• Frameworks: MONAI, PyTorch Lightning
• Task: Multi-class classification
• Loss Function: CrossEntropyLoss

Workflow:

1. Environment Setup: Installs necessary libraries like MONAI and PyTorch Lightning.
2. Data Handling: Downloads and extracts a dataset of DR images. It then dynamically creates file lists for training, validation, and test sets.
3. Data Augmentation: Uses MONAI transforms (LoadImageD, ScaleIntensityD, ResizeD, RandZoomD) to prepare the data for the model.
4. Training: A pl.LightningModule is defined to encapsulate the DenseNet121 model, optimizer, and training/validation/test steps. The training is managed by a PyTorch Lightning Trainer, utilizing callbacks for model checkpointing and early stopping, and logging results to TensorBoard.
5. Inference: After training, the notebook shows how to load the best-saved model checkpoint to perform inference on a single test image and visualize the predicted grade.
6. Monitoring: Includes checks for GPU VRAM usage before and after training.

3. 3D Spleen Segmentation from CT Scans

Notebook: lab_3_Spleen_Segmentation_3D/lab_3_spleen_segmentation_3d.ipynb

This tutorial covers a 3D volumetric segmentation task to delineate the spleen from abdominal CT scans, using the Medical Segmentation Decathlon (MSD) dataset.

Key Components:

• Model: 3D UNet
• Frameworks: MONAI, PyTorch
• Task: Volumetric semantic segmentation
• Loss Function: DiceLoss
• Metric: DiceMetric

Workflow:

1. Data Handling: Downloads and prepares the MSD Spleen dataset. It uses CacheDataset to accelerate data loading during training.
2. Advanced Transforms: A robust pipeline of MONAI transforms is used for pre-processing and augmentation, including spacing correction, intensity scaling, foreground cropping, and balanced patch sampling (RandCropByPosNegLabeld).
3. Training Loop: Implements a standard PyTorch training loop, evaluating the model every two epochs using sliding_window_inference for robust validation on the full-volume images.
4. Evaluation: The model's performance is tracked using the DiceMetric. The best model is saved based on the highest validation Dice score.
5. Analysis: Plots the training loss and validation Dice metric over epochs to visualize the learning progress.
6. Inference: Demonstrates how to load the best model to run inference on validation images and visualize the input, ground truth, and predicted segmentation side-by-side. It also includes a section for running inference on a separate test set.

4. MONAI Label with Self-Trained Model

Notebook: lab_4_MONAI_LABEL_selfmodel_OHIF/lab_4_monailabel_radiology_self_model_OHIF.ipynb

This notebook shows how to integrate your own custom-trained model (from Lab 3) into the MONAI Label ecosystem.

Key Components:

• App: MONAI Label Radiology App (Customized)
• Model: Custom trained 3D UNet (from Lab 3)
• Integration: Custom Inference Class

Workflow:

1. Setup: Prepare the MONAI Label environment and download the radiology app.
2. Customization: Create a custom configuration file (trained_spleen_segmentation.py) to define the inference logic for the self-trained model.
3. Deployment: Start the MONAI Label server, passing the custom configuration and the path to the best checkpoint saved in Lab 3.
4. Validation: Access the OHIF viewer to verify that the self-trained model appears in the plugin and can successfully segment new images.

5. 3D Brain Tumor Segmentation with Swin UNETR

Notebook: lab_5_Brain_Tumor_Segmentation_3D/lab_5_swin_unetr_brats21_segmentation_3d.ipynb

This notebook tackles the challenging task of multi-class brain tumor segmentation from 3D multi-modal MRI scans, using data from the BraTS 21 challenge. It segments three tumor sub-regions: Enhancing Tumor (ET), Tumor Core (TC), and Whole Tumor (WT).

Key Components:

• Model: SwinUNETR (Swin Transformers for Semantic Segmentation)
• Frameworks: MONAI, PyTorch
• Task: Multi-class volumetric semantic segmentation
• Loss Function: DiceLoss
• Metric: DiceMetric

Workflow:

1. Data Management: Downloads the BraTS 21 dataset and uses a JSON file to organize the data into training and validation folds.
2. Specialized Transforms: The transform pipeline is tailored for the BraTS dataset, featuring ConvertToMultiChannelBasedOnBratsClassesd to handle the specific label structure (NCR, ED, ET).
3. State-of-the-Art Model: Utilizes the SwinUNETR model, a powerful transformer-based architecture for 3D medical image segmentation. Gradient checkpointing is enabled for memory-efficient training.
4. Structured Training: A comprehensive training script is built with helper classes (AverageMeter) and functions (train_epoch, val_epoch, trainer) to manage the training and validation process.
5. Inference and Evaluation: Uses sliding_window_inference for validation. The performance is evaluated for each tumor sub-region (TC, WT, ET) individually.
6. Results Visualization: Plots the training loss and the Dice scores for each of the three sub-regions, providing a detailed view of the model's performance.