Dose reduction for medical imaging using deep convolutional neural networks

1. A method of reducing radiotracer dose for radiology imaging modalities and nuclear medicine applications, comprising: using a convolutional network to generate a standard-dose nuclear medicine image from a low-dose nuclear medicine image, wherein said convolutional network comprises N convolution neural network (CNN) stages, wherein each said CNN stage comprises M convolution layers having K x K kernels, wherein said convolutional network further comprises an encoder-decoderstructure having symmetry concatenate connections between corresponding said CNN stages;

wherein said convolutional network implements downsampling using pooling and up-sampling using bilinear interpolation between said stages,

wherein said convolutional network extracts multi-scale and high-level features from said low-dose image to simulate the standard-dose image; and

wherein said convolutional network implements uses concatenate connections to preserve local information and resolution of said standard dose image,

wherein said standard-dose image comprises a dose reduction factor (DRF) equal to 1 of a radio tracer in a patient, wherein said low-dose image comprises a DRF equal to at least4 of said radio tracer in said patient.

2. The method according to claim 1 , wherein said DRF is in a range of 4 to 200.

3. The method according to claim 1 , wherein said standard-dose nuclear medicine image is generated from said low-dose nuclear medicine image and corresponding multi-contrast MR images as multi-modality inputs.

4. The method according to claim 1 , wherein said nuclear medicine image is generated using methods selected from the group consisting of CT, PET, PET/CT, PET/MIR, SPECT, and other nuclear medicine imaging methods.

5. The method according to claim 1 , wherein a signal-to-noise-ratio (SNR) in said low-dose nuclear medicine image is increased using an encoder-decoder residual deep network with concatenate skip connections, wherein said skip connections comprise a residual connection from an input to an output of said method, or concatenating connections between corresponding encoder and decoder layers.

6. The method according to claim 1 , wherein said low-dose nuclear medicine image further comprises a combination of multiple slices and multiple contrast images as input.

7. The method according to claim 6 , wherein said combination of said multiple slices and said multiple contrast images are selected from the group consisting of T1w MR images, T2w MR images, FLAIR MR images, Diffusion MR images, Perfusion MRI images, susceptibility MR images, MR based Attenuation Correction Maps, MR water-fat images, CT images, and CT based Attenuation Correction Maps, wherein said Perfusion MRI images comprise Arterial Spin Labeling sequences.

8. The method according to claim 1 further comprising an algorithm to determine how many input slices and which input contrasts are contributing the most to the method, wherein said algorithm adaptively decides how many said input slices and said input contrasts to use.

9. The method according to claim 1 , wherein mixed cost functions selected from the group consisting of L1/Mean-absolute-error, structural similarity loss, and adaptive trained loss are used, where said adaptive trained loss comprises generative adversarial network loss and perceptual loss function using network models.

10. A system of generating high-quality images for radiology imaging modalities and nuclear medicine applications from low-radiation-dose samples comprising:

a) using a medical imager for taking multiple slices of low-radiation-dose images, or low-radiation-dose images and multi-contrast images acquired together, as a stacking of multiple 2 dimensional images or 3 dimensional images as a system input;

b) applying a deep network-based regression task to said input images, wherein said deep network-based regression task comprises;

i. N convolution neural network (CNN) stages, wherein each said CNN stage comprises M convolution layers having K×K kernels, wherein said CNN comprises an encoder-decoder structure having symmetry concatenate connections between corresponding said CNN stages;

ii. an encoder-decoder residual deep network with concatenate skip connections, wherein said skip connections comprise a residual connection from an input image to an output image; and

iii. outputting radiology or nuclear medicine images having an image quality as a standard-radiation-dose image, wherein said image quality comprises a resolution, a contrast, and a signal-to-noise-ratio that are improved from low-radiation-dose inputs.