Research

Modular Material Decomposition from Photon Counting CT

Publishing soon

Currently, I am working on an efficient and modular framework for photon counting CT (PCCT) reconstruction and material decomposition. PCCT can dramatically improve clinical CT by providing excellent quality reconstructions at the reduced radiation exposure!

Texture Matching GAN (TMGAN)

Paper: coming soon Patent: 63/410,486

To produce desired texture while denoising and/or sharpening, we propose a generative model. To alleviate the risk of hallucination in medical CT, TMGAN separates anatomy from texture using a Siamese network.

Idea: Match the distribution of generated texture (only)

Results: Quantitative and Qualitative

Bias-Reducing Cost Function

PDF Poster IEEE Xplore Patent Contact for code

To preserve texture we present a novel approach to designing a cost function that penalizes variance and bias differently, as opposed to equal penalties in the mean squared error (MSE).

Idea: Bias weighted MSE realized with a Siamese network

Results: Clincal Exam

Noise Preserving Sharpening Filter (NPSF)

PDF Poster IEEE Xplore Patent in progress Contact for code

To preserve texture while sharpening, we preserve noise energy and texture from the input by adding appropriately scaled noise while training. NPSF has been favored clinically!

NPSF concept: Appropriate Training Pairs to Reduce Bias