Research · 2022 · Research at KJ Somaiya College of Engineering · IEEE ICAST 2022
MRI Denoising for Brain Tumour Detection
Can a model diagnose through the noise?
A two-level U-Net that strips Gaussian noise from brain MRI scans, lifting PSNR from 11.90 to 30.96 and raising tumour classification accuracy by 23% on average across six CNNs. Published at IEEE ICAST 2022.
Problem
Thermal noise and motion leave brain MRI scans grainy enough to change a tumour diagnosis. The scan has to be cleaned before a classifier sees it.
Approach
- 253 brain MRI scans from a public Kaggle set, resized to 320×320 and augmented six-fold to 1,518 images.
- A noisy copy of every image made with Gaussian noise (mean 0.25), matched by eye to clinical noise.
- A U-Net with two encoder-decoder pairs (64 and 128 filters, dropout 0.2), compared against a residual U-Net and four classical filters.
- Six CNN classifiers (DenseNet201, VGG16, InceptionV3, EfficientNetB7, ResNet101 and a custom CNN) trained on noisy and on denoised scans.
Stack
- Python
- TensorFlow
- Keras
- OpenCV
Outcome
- PSNR raised from 11.90 to 30.96
- Tumour classification accuracy up 23.02% on average after denoising
- Shallower U-Nets denoised better: PSNR fell from 30.96 at depth 2 to 17.87 at depth 6