Deep diffusion models for anomaly detection

Keywords

Deep Learning, Weakly supervised segmentation, Diffusion models


Scientific context

Deep learning is now an established approach for medical image processing, providing
state of the art results for image segmentation. However, best results are obtained with supervised learning,
requiring a large annotated dataset. Especially for medical image segmentation, annotation is a tedious
and time consuming task, requiring experienced clinician as annotator. In a previous work, we propose to
use constrained learning with attribution maps to segment tumor or multiple sclerosis (MS) lesions in a
weakly supervised way. Deep unsupervised learning is often based on generative models such as generative
adversarial network (GAN) or variational autoencoder (VAE) but these models has strong limitations: GAN
are notoriously difficult to train and VAE produces blurry images. Recently, diffusion model [Song et al] have
shown there ability to generate images that can be highly resolved and trained in a stable manner. Some
authors already proposed to use them for anomaly detection on medical images [Wolleb et al].

Objective

The objective of the internship would be to propose a new weakly supervised anomaly detection
method based on diffusion models. A work program could be: 1/ deeply understand the mechanism of
diffusion models, 2/ reproduce the results of [Wolleb et al] and apply them to different data and different
pathology (MS), 3/ combine the methods of [Wargnier et al] and [Wolleb et al] and evaluate the performance
of the combined method. This work program can evolve based on the results obtained during the internship
and new ideas of the intern.

Data

Several datasets are already available for use in the lab.

Application

The successful candidate is expected to be autonomous, to show strong motivation and in-
terest in multidisciplinary research (image processing and machine learning for medical applications), to be
highly proficient in python (pytorch is a plus) and to have a background in either
• Image analysis
• deep learning
• or applied mathematics
Interested applicants are required to send a cover letter, resume and any other relevant documents (ref-
erence letter, recent transcripts of marks,...) to: michael.sdika[at]creatis.insa-lyon.fr, valentine.
wargnier[at]creatis.insa-lyon.fr and thomas.grenier[at]creatis.insa-lyon.fr.

Bibliography

  1. J. Song et al, Denoising Diffusion Implicit Models, ICLR 2021
  2. J. Wolleb et al, Diffusion Models for Medical Anomaly Detection, MICCAI 2022,
  3. V. Wargnier Dauchelle, A Weakly-Supervised Gradient Attribution Constraint for Interpretable Classification and Anomaly Detection, et al, IEEE TMI (submitted)