Segmentation of biomedical images
One of the fields of application of Artificial Intelligence in medicine is the segmentation of biomedical images. Image segmentation is an image processing that consists in grouping pixels of an image according to some of their characteristics.
Great challenges to advance research faster
The website https://grand-challenge.org/challenges lists 230 challenges organized since 2008 in the field of medical image analysis. A great challenge is a contest to create models created from the same data set. The team of researchers whose model is the most efficient wins the grand challenge.
The big challenges are often organized by medical universities, hospitals, research laboratories, private companies such as Google, Nvidia, Intel, etc… They provide researchers with reliable data and help create a competitive environment that advances research faster.
U-Net for biomedical image segmentation
U-Net is a convolutional network architecture (CNN) allowing fast and accurate segmentation of biomedical images (microscopes and X-rays). U-Net won the “Grand Challenge for Computer-Automated Detection of Caries in Bitewing Radiography at ISBI 2015”.
U-Net has the advantage of running on medium to small data sets that match the medical data characteristic (about 30 images is sufficient) and is easy to train.
It detects the context and the objects ; then precisely locates the contour of the detected objects. U-Net is perfectly adapted to the detection of tumors.
Optimization of U-Net performance
Artificial Intelligence research consists of optimizing the performance of neural networks while reducing the power required for calculations. Since 2015, researchers have been working to optimize the structure of U-net to improve its performances.
- UNet++enhances segmentation accuracy by including a dense block and convolution layers between encoder and decoder.
- Attention U-Net allows you to add attention mechanisms (extraction of areas of interest from the image) that improve model performance (sensitivity and accuracy) without increasing the required computational capacity.
- theMultiResUNet is an enhanced version of the standard U-Net architecture, developed in 2020 by Samsung and the Bangladesh University of Engineering and Technology. MultiResUNet has outperformed U-Net on five datasets.
- In February 2021, the TransUNet model was introduced as a hybrid version of U-Net and Transformers that can exploit the capabilities of both architectures. TransUNet is inspired by Vision Transformer (ViT).
U-Net to fight against brain tumors
Brain tumors develop within the skull. They may be metastases, secondary to another cancer initially developed outside the brain (most often breast, lung, kidney or skin cancer) or primary brain tumours, originating from cells normally present in the central nervous system.
MRI scans of the brain are essential if a tumor is suspected, and certain MRI sequences provide additional information about the tumor’s “behavior”: the size of vessels, cell density, and metabolic impact.
The following example from the course Coursera Specialization AI for Medicine shows how to implement a U-net with Keras to automatically segment tumor regions in the brain.
Choice of cost function: Soft Dice Loss
Creation and training of the Unet_model_3d model with the Soft Dice Loss cost function
Once trained the U-Net model is able to make predictions
image, label = load_case(DATA_DIR + "imagesTr/BRATS_003.nii.gz", DATA_DIR + "labelsTr/BRATS_003.nii.gz")
pred = util.predict_and_viz(image, label, model, .5, loc=(130, 130, 77))
Sensitivity and Specificity calculations are used to evaluate the model
To see the complete Brain Tumor Auto-Segmentation for Magnetic Resonance Imaging (MRI) code
Appendix
Implementation of U-Net with Keras
