The tutorial demonstrates how to train a cell segmentation model using the MONAI framework and the Segment Anything Model (SAM) on the Cellpose dataset.
In Summary the tutorial covers the following:
- Initialization of the CellSamWrapper model with pre-trained SAM weights
- Creation of data lists for training, validation, and testing
- Definition of data transforms for training and validation
- Setup of datasets and dataloaders with MONAI
- Implementation of the training loop, including:
- Loss function (CellLoss)
- Accuracy function (CellAcc)
- Optimizer (SGD)
- Mixed precision training with GradScaler
- Sliding window inference via MONAI
- Visualization of training loss, validation loss, and validation accuracy
- Inference on a single validation image
- Visualization of input image, ground truth, and model prediction
For a more elaborate experience we encourage you to take a look at the VISTA-2DMONAI bundle.