SELMA3D-XunDJ-final_model
About
- 3D brain microscopy (3D microscopy image of the brain)
- Biological brain structure (Segmentation of a biological structure in the brain with 0-background and 1-anatomical structure)
Challenge Performance
| Date | Challenge | Phase | Rank |
|---|---|---|---|
| Sept. 24, 2024 | SELMA3D | Preliminary Test Phase | 11 |
| Sept. 25, 2024 | SELMA3D | Final Test Phase | 3 |
Model Facts
Summary
Final version for SELMA3D
Mechanism
Pretraining Firstly, we use the unannotated dataset for pretraining. We use a classic network architecture Swin UNETR which is widely used in 3D medical image analysis. Before training the model, we implemented several preprocessing steps. (1) We scaled each 2D plane image down to 1/3 of its original width and height to optimize memory usage and converted the ‘tif’ images to ‘png’ format for compatibility with the MONAI Python package. (2) We created a sample dictionary, allowing for random sampling of continuous sets of plane images along the z-dimension. (3) A 3D patch was randomly selected from the 3D image, followed by normalization, cropped and padding (if the image was smaller than the required input size). The model input size is configured to 64x64x64. Instead of the conventional inpainting transformation, we adopt a masked self-supervised learning strategy. In contrast to the original inpainting method, our approach directly sets certain regions to zero, elevating the pre-training challenge and encouraging the model to learn more robust features. The batch size was set as 4 and the epoch was 100.
Finetuning For fine-tuning, we utilized the annotated data, applying several preprocessing techniques: normalization, random cropping, padding, random flipping, random rotation, and random intensity variation. The batch size was set to 4, with training conducted 2000 epochs. We employed the DiceMetric from the MONAI Python package to assess performance on validation images, and applied a threshold to convert the model outputs into binary format. The best performing and final checkpoints on the validation set during training was saved for submission.
Validation and Performance
Uses and Directions
This algorithm was developed for research purposes only.
Warnings
Common Error Messages
Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3