Brain Tumor Detection with Computer Vision

This work aims to identify the location of brain tumors in MRI images through the use of object-detection.

These are personal contributions and continuations of a group project for CS4100. You can find the project report here.

Datasets

The datasets used in this project are available on Kaggle:

• Brain Tumor Image Dataset: Semantic Segmentation

Find the data exploration for this dataset here.

Problem Formulation

Single Object Detection

Directly predict the coordinates of the ground truth bounding box (center of the box and its width, height).

Benefit: simple and naive approach

Drawbacks: relies on the assumption that there is exactly 1 tumor bounding box per image

Anchor-based Object Detection

Pre-define anchors boxes of various sizes and aspect ratios. For each anchor box, make a classification prediction (is there an object in this anchor?) and a regression prediction (how to make the anchor box overlay the true object bounding box).

Benefits: can handle cases when there are no tumors or many tumors in a single image

Drawbacks: more difficult to train due to spacity of the labels; more complicated model architecture.

Setup Instructions

Step 1: Install Dependencies

Install the necessary Python packages using the requirements.txt file:

pip install -r requirements.txt

Alternatively, you can create a conda environment using the following commands:

conda env create -f environment.yaml
conda activate brain_tumor_detection_env

Step 3: Kaggle API Authentication

Follow the instructions to set up your Kaggle API credentials. You can find the Kaggle API authentication instructions in the Kaggle API Documentation.

Step 4: Download Datasets

Refer to the notebooks/downloading_datasets.ipynb notebook for step-by-step instructions on using the Kaggle API to download the datasets required for this project. The datasets will be downloaded to the ./datasets folder, which is configured to be ignored by git.

Running Experiments:

Single Object Detection

Navigate to the main function in src.training.train_single_object_detection and edit the training_configs.

TrainingConfig Arguments:

• dataset_root_dir
• device
• batch_size
• learning_rate
• num_epochs
• optimizer ("Adam" or "SGD")
• image_size
• pretrained_backbone (True or False)
• efficient_net_version: ("b0", ... "b7")
• predictor_hidden_dims (e.g. [[64, 16])
• augmentations: ("rotation", "reflection", and/or "crop")

To run the specified experiment you can use...

python -m src.training.train_single_object_detection

Anchor-Based Object Detection

Some Experimental Results

Single Object Detection

Experimental Configs:

{
  "batch_size": 16,
  "learning_rate": 0.001,
  "num_epochs": 100,
  "device": "mps",
  "dataset_root_dir": "./datasets",
  "augmentations": ["rotation", "reflection", "crop"],
  "optimizer": "Adam",
  "image_size": 256,
  "pretrained_backbone": true,
  "efficient_net_version": "b4",
  "predictor_hidden_dims": [64, 16],
  "save_dir_path": "./experiments/24-05-31-01-09",
  "git commit id": "ff97464a381f489a97e7730756a276446a4a03ac"
}

Experimental Results:

{
  "best_val_iou": 0.6572,
  "best_epoch": 67,
  "test_iou": 0.6420
}

Example Test Predictions: