Using Diffusion Models for Hematopathology Image Synthesis for Medical Diagnostics

Project Overview:

Diffusion models are models that can create high-quality images, videos, or audio by learning to turn random noise into an image. These models have recently enabled highly realistic image generation and have shown strong performance in medical imaging data applications such as histopathology and cytomorphology. Prior research demonstrates that these generative models can improve model performance in environments with limited data by augmenting training datasets with synthetic images. To ensure efficient annotation, only the most relevant objects should be presented to experts for labeling. However, there are currently no standardized tools for selecting relevant objects in multiclass detection tasks. Our work aims to advance this area by experimenting with image synthesis methods for cytomorphology.

CytoDiff: AI-Driven Cytomorphology Image Synthesis for Medical Diagnostics

My Project:

I am partnering with Dr. Umer and Dr. Marr from the Institute of AI for Health (AIH) in Munich, Germany, to create a diffusion model to generate synthetic medical imaging data to improve the performance of classification models and expand currently limited annotated datasets. With the incorporation of these synthetic images, hospitals could train more accurate machine learning models to recognize diseases from microscope images of blood or tissue samples. We are focusing on generating images of white blood cells and bone marrow smears. The classification of individual white blood cells is key to diagnosing hematological malignancies such as leukemia. We are working on building upon the CytoDiff paper's framework for their model. However, instead of training the model on the image space, we are going to train the model on embeddings (vectors that represent features of an image) using a foundation model created by the Marr lab to extract them. From there, we will test the generated samples using a variety of evaluation metrics and test them with classifiers. These classifiers will be trained with both synthetic and real data to see if their ability to accurately identify a white blood cell type improves.

Here is a link to the lab's page: https://www.helmholtz-munich.de/en/aih/research-groups/marr-group

Page updated

Report abuse