Brain Cancer Modelling URA

• Code from https://github.com/KurtLabUW/brats2023_updated?tab=readme-ov-file
• Using BraTS 2023 Data from Synapse https://www.synapse.org/Synapse:syn51156910/files/

Reproducing Results

This repository contains pipelines for 3D tumor segmentation using U-Net and boundary refinement using the Segment Anything Model 3 (SAM 3). Below are instructions to run the models from scratch.

Prerequisites

Ensure you have installed the required dependencies:

pip install -r brats2023_updated/requirements.txt

Note: Make sure your dataset and checkpoint paths in the __main__ blocks of the respective training/inference scripts are pointing to the correct locations on your local machine.

U-Net Model

1. Training To train the base 3D U-Net model on the BraTS dataset, execute the following from the root directory:

python -m brats2023_updated.model_routines.train_with_val

2. Inference Once trained, to run inference (predicting on validation/test data) with the U-Net model, run:

python -m brats2023_updated.model_routines.infer

SAM 3 Model (Fine-Tuning & Inference)

The SAM 3 pipeline utilizes a fine-tuned SAM 3 model for slice-by-slice boundary refinement.

Option 1: Complete Pipeline Script You can run the full SAM 3 pipeline (training followed by inference) using the provided bash script:

./brats2023_updated/sam/run_sam3_pipeline.sh

Option 2: Manual Execution

1. Training (LoRA Fine-tuning) To fine-tune the SAM 3 model using LoRA adapters:

python -m brats2023_updated.sam.train_sam3

2. Inference To evaluate the fine-tuned SAM 3 model (this uses bounding boxes to prompt the model):

python -m brats2023_updated.sam.infer_sam3