Denoising medical images by learning sparse image representations with a deep unfolding approach using scan specific metadata
The present embodiments relate to denoising medical images. By way of introduction, the present embodiments described below include apparatuses and methods for machine learning sparse image representations with deep unfolding and deploying the machine learnt network to denoise medical images. Iterative thresholding is performed using a deep neural network by training each layer of the network as an iteration of an iterative shrinkage algorithm. The deep neural network is randomly initialized and trained independently with a patch-based approach to learn sparse image representations for denoising image data. The different layers of the deep neural network are unfolded into a feed-forward network trained end-to-end.
1. A method for denoising medical images in a computed tomography (CT) system, the method comprising:
scanning, by the CT system, a patient to generate CT image data;
denoising, by an image processor, the CT image data by applying a deep-learnt multiscale network of filters decomposing the CT image data into sparse image representations at multiple scales, the deep-learnt multiscale network of filters comprising a cascade of trained sparse denoising autoencoders, wherein lower levels of the deep-learnt multiscale network apply learnt filters to image data recursively downsampled from the CT image data and resulting denoised image data at each scale is upsampled back to full resolution and summed to obtain a final denoised image CT data set, and wherein inputs to the sparse denoising autoencoders were a concatenation of CT pixel data and metadata of the CT system; and
displaying, by a display, an image rendered from the final denoised CT image data set.
2. The method of claim 1 wherein the filters trained by the cascade of sparse denoising autoencoders were randomly initialized and trained jointly and end-to-end to optimize a quality of the denoised image data.
3. The method of claim 1 wherein filter weights of the filters are initialized with a random zero-mean Gaussian distribution.
4. The method of claim 1 wherein the filters trained by the cascade of sparse denoising autoencoders were trained with image data patches.
5. The method of claim 4 wherein denoising is repeated for patches of the CT image data at each scale.
6. The method of claim 5 wherein denoising learnt filter maps the patch of the CT image data to sparse image representations, wherein the sparse image representations are thresholded to remove noise.
7. The method of claim 6 wherein a threshold for the thresholding is a function of the metadata.
8. The method of claim 1 wherein the sparse denoising autoencoders were trained with a noise model directly related to a radiation dose of the CT system.
9. The method of claim 1 wherein the metadata of the CT system comprises a dose, a reconstruction resolution and a reconstruction kernel.