Generative adversarial network-based lossless image compression model for cross-sectional imaging
A novel medical image generation approach synthesizes thin-cut computerized tomography (CT) images from thick-cut CT ones as inputs. First thick-cut CT images are obtained by maximizing the pixel-wise intensity of five or more than five continuous thin-cut ones after image registration. Second, the obtained thick-cut CT images are fed into a generator block which adopts an encoder-decoder architecture, where each thick-cut image is encoded into low-dimensional embedding space before decoding into multiple thin-cut ones. Third, a discriminator focuses on distinguishing original real thin-cut images from synthetic thin-cut images. An adversarial mechanism between the generator and discriminator causes the discriminator's output to provide an effective gradient update of the network parameters for the generator to increasingly improve the generator's ability to synthesize higher-quality thin-cut images and in turn promotes the discriminator's discriminating capability.
1 . An end-to-end thick-to-thin cut generation computer framework for CT imaging to recover near-original thin radiological images from compressed low-quality thick radiological images, comprising:
a generator module responsible for generating synthetic high-quality thin-cut images from compressed cross-sectional low-quality thick radiological images;
a discriminator module that receives real and synthetic thin CT images and which distinguishes real uncompressed thin-cut images from synthetic thin-cut images; and
an adversarial mechanism between the generator and discriminator that causes an output of the discriminator to provide an effective gradient update of the network parameters for the generator to increasingly improve the generator's ability to synthesize higher-quality thin-cut images and in turn promotes the discriminator's discriminating capability,
whereby the generator module's quality is improved by a cost function that combines pixel-wise reconstruction loss, content loss and depth estimation loss.
2 . The thin cut generation computer framework according to claim 1 wherein pixel-wise reconstruction loss allows for the preservation of fine details while resulting in blurring images, the content loss allows the preservation of high-level semantic information, and the depth estimation loss is used to quantify the differences between the real thick-cut images and the generated synthetic thick-cut images at the pixel level and semantic level.
3 . The thin cut generation computer framework according to claim 1 wherein first thick-cut CT images are obtained by maximizing the pixel-wise intensity of multiple continuous thin-cut images after image registration and then the obtained thick CT images are fed into the generator to synthesize thin CT images.
4 . The thin cut generation computer framework according to claim 1 wherein the multiple continuous synthetic thin-cut images are five continuous thin-cut images.
5 . The thin cut generation computer framework according to claim 1 wherein the generator includes a first stage composed of three convolutional layers adapted to learn high-level semantic features, a second stage composed of nine continuous residual blocks adapted to learn discriminative features, and a third stage composed of three de-convolutional layers responsible for recovering fine details.
6 . The thin cut generation computer framework according to claim 1 further including a feature extractor module that enforces characteristics on the synthetic thin CT images that are as close as possible to those of real thin CT in order to mitigate undesirable blurring generation by quantifying divergence of high-level semantic information between real thick-cut cross-sectional images and synthetic thick-cut images derived from synthetic thin-cut cross-sectional images.