Medical imaging — MRI denoising and gene expression from pathology

At a glance

• Speckle noise hides diagnostic detail in MRI; gene-expression assays are slow and costly, while pathology images are fast and cheap.
• Speckle-denoising models for MRI, and a thesis predicting gene expression from gigapixel pathology images.
• Completed: MRI project 2020 to 2021; MSc thesis finished September 2025, with a paper in preparation.

Highlights

• Speckle-noise reduction for MRI

  Denoising models aimed at the speckle noise that destroys diagnostic detail in MRI and related modalities.

• Gene expression from pathology images

  The thesis tests whether whole-slide pathology images can stand in for a slow, costly gene-expression assay.

• Vision foundation models

  A vision foundation model processes the gigapixel whole-slide image as the first step of the pathology track.

• Graph neural networks

  A graph neural network builds on the vision model's output on the way to the gene-expression prediction.

• Adapted language-model techniques

  Techniques from large language models, adapted for this task, combined with statistical and probabilistic methods.

• Aiming for low compute and latency

  The thesis aimed to keep processing and latency as low as possible, a hard constraint for gigapixel inputs.

The challenge

Speckle noise destroys diagnostic detail in MRI. The first project set out to reduce it, in MRI and related imaging modalities.

The second problem is economic as much as scientific. Gene-expression assays are slow, expensive and logistically painful, while whole-slide pathology images are fast and cheap to obtain. Those images are gigapixel-scale, and the thesis set a hard constraint on compute and latency.

Architecture

Two research tracks. For MRI, speckle-denoising models. For pathology, a vision foundation model processes the whole-slide image and a graph neural network builds on its output, while large-language-model techniques adapted for the task also feed the gene-expression prediction. The approach also uses statistical and probabilistic methods.

What I built

The thesis asks whether whole-slide pathology images can stand in for a gene-expression assay. It aims to recover molecular-biology signal directly from gigapixel images, and to keep processing and latency as low as possible while doing so.

• MRI: speckle-noise reduction in MRI and related imaging modalities.

• Pathology: gene expression predicted from digital pathology images, combining graph neural networks, vision foundation models, adapted LLM techniques and statistical modelling.

Outcome

The MSc thesis was completed in September 2025 at Isfahan University of Technology, and a paper based on it is in preparation. The MRI speckle-noise project for a private company ran from March 2020 to March 2021.

My role

• Researched speckle-noise reduction in MRI and related imaging modalities for a private company, March 2020 to March 2021.
• Wrote an MSc thesis at Isfahan University of Technology on predicting gene expression from digital pathology images.
• Combined graph neural networks, vision foundation models, adapted LLM techniques and statistical methods in the thesis approach.
• Worked under a hard constraint on compute and latency in the thesis.

Engineering

• Compute and latency as constraints

  The thesis worked under a hard constraint on compute and latency, aiming to keep both as low as possible.

• Designed for gigapixel inputs

  The pathology track works directly on gigapixel whole-slide images.

• Several model families, one prediction

  Graph, vision-foundation and adapted language-model methods feed a single gene-expression prediction, with statistical methods alongside.

Outcomes

• MSc thesis completed in September 2025 at Isfahan University of Technology; a paper is in preparation.
• MRI speckle-noise reduction project completed for a private company, March 2020 to March 2021.
• Thesis: Predicting gene expression (molecular biology) from digital pathology images.