Competitive Advantages
- Provides a fast and lean solution to denoise medical images
- Could improve diagnostic accuracyand thus increased efficiency of the resulting treatment
- Denoising is based on patient’s individual imaging featureswithout need for pretraining or transfer learning
Summary
Medical imaging provides accuracy in diagnosis in a wide variety of specialties which requires the information on these images to be as accurate as possible. Noise artifact which occursduring acquisition or in post-processing inmedical imaging is a significant problem to diagnosticians as it can mimic pathology. Thisinvention develops a computationallylightweight algorithm and provides a rapid and adaptive denoising strategy to handle these artifacts.For this method a PY file is developed which imports libraries for feature extraction and image composition. Any nxn dimensional image with RGB channels can be evaluated by thisalgorithmthe image will beresizedto 256x256x3 pixels and then convertedto a Numpy array of equal size for matrix calculation. Thisimage isthendecomposed into 8x8 patches in the 2D matrix space and a Dictionary of x features to be extracted over i iterations, where both variables x, iare integer values. A greedy pursuit algorithm known as Orthogonal Matching Pursuit (OMP) is then used to approximate the denoised patches from a tunable parameter. The denoised patches are then reassembled and the denoised image as a whole is shown along with the noise subtracted.
Figure 1illustrates a noisy Computed tomography pulmonary angiography(CTPA)image, denoised image, and the noise differenceafter a single sparse atom requirement (computational time on NVIDIA V100 GPU: 0.7 seconds). Figure 2a and 2b demonstrate the same denoising algorithm clarifying the presence of a multiple sclerosis lesion versus artifact in the thoracic spine.
Desired Partnerships