UCLA researchers in the Department of Electrical and Computer Engineering have developed an algorithm for hallucination-free resolution enhanced brain MRI images with better quality and superior computational time compared to the current state of the art.
Medical imaging is a widely used diagnostic tool, yet faces challenges with low spatial resolution, limited dynamic range and low contrast. These challenges have fueled interest in enhancing medical images using digital post processing. New possibilities for improving the quality of medical images have included machine learning techniques such as deep neural networks. However, these techniques require large training data sets and significant computational resources which may not always be available in medical imaging. Moreover, neural network-derived hallucination image details may lead to false diagnostics. Computationally efficient image upscaling methods such as Rapid and Accurate Image Super Resolution (RAISR), recently introduced by Google for restoration of natural images, demonstrate high-quality restoration while being about 2 orders of magnitude faster than other techniques that produce similar quality. Thus, these types of architectures have potential for improving the quality of medical images where image sizes can be very large and computational resources may be limited.
UCLA researchers have developed an algorithm for hallucination-free resolution enhanced brain MRI images. This algorithm is useful for overcoming tradeoffs between image resolution, brightness and contrast detectability in MRI imaging. A locally adaptive learned filtering technique is used to improve the resolution of brain MRI images. This algorithm achieves superior performance in visually impaired images (with a peak signal-to-noise ratio about 0.3dB higher than RAISR algorithm) as well as low computational complexity in locally adaptive regression-based learning. Moreover, the images have improved sharpness and no hallucination. Applications to other medical imaging modalities such as CT/PET and digital X-ray is expected, where the technique can provide good quality images at a lower radiation dose. The algorithm can also be extended for de-noising, dynamic range improvements, image compression and color reconstruction.
medical imaging, brain imaging, MRI, locally adaptive learned filtering, deep neural networks, neural network algorithm, phase stretch anchored regression, phase stretch transform