Paul Friedrich

I am a third year PhD student at the Center for medical Image Analysis and Navigation (CIAN) of the University of Basel. My work is supervised by Philippe Cattin and Florian Thieringer. I hold a bachelor's and a master's degree in biomedical engineering, both of which I finished best in class.

Email  /  Google Scholar  /  Github  /  LinkedIn

I'm interested in medical image analysis, computer vision, 3D deep learning and generative models. My current research is part of the MIRACLE II project and focuses on conditional shape generation and manipulation tasks as well as (un)conditional 3D medical image generation. I'm particularly interested in different types of 3D data representations, including voxel grids, point clouds, meshes, triplanes, or neural implicit representations.

A review of deep generative models (VAEs, GANs and DDMs) for 3D medical image synthesis. This book chapter covers fundamental principles, recent advaces, as well as strength and weaknesses of different deep generative models for 3D medical image synthesis and examines their application for relevant downstream tasks. The chapter also reviews commonly used evaluation metrics for assessing image fidelity, diversity, utility and privacy.

An implicit neural representation (INR), to predict 2D- and 3D MR images of neonatal brains at varying time points.

A review on different diffusion models for 3D healthy brain tissue inpainting.

A Bernoulli Diffusion Model operating on a binary latent representation of the data to effectively solve an unsupervised anomaly detection task.

WDM is a framework for high-resolution medical images synthesis. We propose a simple yet effective way of scaling 3D diffusion models to high resolutions (256 x 256 x 256 on a single 40 GB GPU) by applying Discrete Wavelet Transform for spatial dimensionality reduction.

MedShapeNet contains over 100,000 medical shapes, including bones, organs, vessels, muscles, etc., as well as surgical instruments. You can search, display them in 3D and download the individual shapes by using our shape search engine.

Following the recent success of diffusion models, we propose a novel approach for automatic implant generation based on a combination of 3D point cloud diffusion models and voxelization networks. Due to the non-deterministic sampling process in our diffusion model, we can propose an ensemble of different implants per defect, from which the physicians can choose the most suitable one.

• Conference reviewer: MICCAI, DGM4MICCAI
• Others: Member of the Integrity Comission @ DBE University of Basel

I'm currently having an open position for a master's thesis project on Improved Point Cloud Diffusion Models for Automatic Implant Generation. Drop me a mail in case you are interrested.

template adapted from this website