Brain Lesion Model

Introduction

• We created a conditional adversarial network model that identifies, analyzes, and isolates brain lesions from MRI scans
• Brain lesions simply mean any abnormal brain tissues that appear as bright spots on the brain in MRI scans
• The goal of this project is to increase doctors' diagnosis accuracies relating to identifying unhealthy brain tissues
• This increases the likelihood of successfully identifying and treating brain abnormalities while preventing unnecessary brain surgeries if the tissue was actually healthy

Training and testing the model

• Obtain MRI scans of brain scans and save them in a folder named 'MRIscans'
  • See 'MRIscans' directory for MRI image examples
• Manually isolate the brain lesions from the images and save them in a folder named 'lesions'
  • See 'lesions' directory for isolated brain lesion examples
• Run ./train.sh to train the custom model
• Run ./test.sh to test the custom model
• Open processed_test/index.html in your default browser to view the brain scan, model's guess, actual isolated lesion, and the accuracy measured against the SSIM index between the guess and the true lesion

Model Results

Sources and Dependencies

• Utilized Materialize to design the displayed HTML results
  • Materialize is a design language created by Google for constructing elegant user experiences
• Modified the code of pix2pix, a tensorflow model for images
  • Altered the splitting, training, and testing portion of the code to tailor the model to learn brain lesion segmentations
  • Fine-tuned the parameters of the model to suit our custom MRI scans
  • Model outputs a file in processed/index.html which displays the final test results
    • Results display brain scan, model's guess of where the lesion is, and the actual isolated lesion
• Incorporated pyssim to test the Structural Similarity Image Metric (SSIM) between the model's guess and the ground truth
  • ./test.sh automatically computes the accuracy between all guesses and actual lesions
  • Once computed, the script updates processed_test/index.html with the corresponding accuracies