Image Projects DIGIT/IMA
Master of Computer Science, Sorbonne Université

Supervisor: mailto:Philippe.Andrey@inrae.fr
Number of students: 2

Project details
(supervision may be done by visioconference if needed)

Supervisor: mailto:Fabrice.Minoletti@sorbonne-universite.fr
Number of students: 2

Supervisor: mailto:Clara.Bremond_martin@sorbonne-universite.fr
Number of students: 2

Project details

Supervisor: mailto:Isabelle.Bloch@lip6.fr
Number of students: 1

This project will be in pair with a medical student.

The diagnosis of cardiovascular deseases benefits from images acquired using magnetic resonance imaging (MRI). Among the parameters of medical interest, the volumes of the left ventricule and of the myocardium and their evolution over the cardiac cycle are of prime importance.

These parameters can be automatically computed from the segmentation of these structures in MRI.

The aim of this project is to investigate image segmentation methods dedicated to this task. Simple methods based on thresholding, k-means clustering and region growing will be developed first, before moving to more sophisticated methods. Extensions to 3D and to time sequences will also be tested, for instance starting from the segmentation at time \(t\) to initialize a segmentation at time \(t+1\).

Supervisors: mailto:Laurent.Beaudoin@epita.fr, loica.avanthey@epita.fr
Number of students: 2

Project details

Supervisors: mailto:Laurent.Beaudoin@epita.fr, loica.avanthey@epita.fr
Number of students: 2

Project details

Supervisor: mailto:Thibaut.Lagache@pasteur.fr
Number of students: 2

In biological imaging, fluorescence microscopy is the gold-standard technique for analyzing the spatio-temporal dynamics of biological processes. When analyzing cells or small organisms such as bacteria, the automatic delineation of the contours of observed objects is often required over time to track the morphological changes of the cell/organism. This project will consist of exploring some standard methods (intensity thresholding and morphological analysis, active contours…) to automatically delineate small fluorescent organisms (small freshwater cnidarian Hydra that are cultured in the lab) and track their behavior.

Supervisor: mailto:Fabien.Verite@isir.upmc.fr
Number of students: 2+2

Project details

Supervisor: mailto:Garance.Lucas@lip6.fr
Number of students: 2

Colorectal cancer (CRC) is a leading cause of cancer deaths worldwide. The standard medical procedure for screening and treating colorectal disease is a colonoscopy.

Histopathological characterization of colorectal polyps is crucial for determining the risk of colorectal cancer and future rates of surveillance for patients. By effectively examining the colonoscopic images to identify precancerous polyps early and remove them before they become cancerous, CRC mortality can be lowered significantly.

Manual examination of colonoscopy images for the detection of precancerous polyps is time-consuming, tedious, and prone to human error. This project aims to develop a deep learning method and image preprocessing techniques for classifying colonoscopic images to assist gastroenterologists in diagnosis.

Supervisor: mailto:Isabelle.Bloch@lip6.fr
Number of students: 2

Opening and closing are morphological filters to remove bright or dark small structures. When the structures are lines, specific shapes of filters have to be designed.

The traditional morphological filter used for the analysis of linear structures in images is the union of openings (or the intersection of closings) by linear segments.

However structures in images are rarely strictly straight, and as a result a more flexible approach is needed, based on constrained paths, i.e. discrete, one-pixel thick successions of pixels oriented in a particular direction, but in general forming curved lines rather than perfectly straight lines. This is called path opening or closing.

The aim of this project is to implement such filters in an efficient way and to test them on various images, in particular medical images.

References:
H. Heijmans, M. Buckley, H. Talbot, Path Openings and Closings, Journal of Mathematical Imaging and Vision 22: 107–119, 2005
H. Talbot, B. Appleton, Efficient complete and incomplete path openings and closings, Image and Vision Computing, 25, 416–425, 2007

Supervisor: mailto:Dominique.Bereziat@lip6.fr
Number of students: 2

The Kinect is a data acquisition device developed by Microsoft. It combines an RGB camera and a depth sensor sensor, named D in the following. A depth map encode for each pixel the distance between the sensor and the objects of the acquisition scene.

The aim of this project is to enrich the information given to a model with depth maps and evaluate the impact on the performance. Several strategies are possible, for instance but other are possible:

• re-training the model using the new data (D or RGB+D), this approach is expensive in terms of computing resources
• add an input layer that merges RGB and D and only train this part of the model only this part
• reduce the dimensionality of the RGB+D input to 3 channels, using ACP, for example, before feeding the model. This last approach will serve as our baseline.

Finally, we will choose a Yolo type model which allows us to track in real time objects detected in a video.

There are datasets that combine RGB images and depth maps (e.g. NYU Depth Dataset, SUN RGB-D, …) that can be used for the project. We also have two Kinect to create a demonstrator.

Supervisor: mailto:Dominique.Bereziat@lip6.fr
Number of students: 2

Nowadays, neural networks have become the state-of-the-art application in many tasks, and in classification task particularly. However these models are notoriously inefficient to understand uncertainty, notably missing data are not correctly handled.

In this project, we propose to take into account uncertainty to improve the performance of a classification model. We propose a very simple method: from an annotated dataset, we mask several regions (randomly chosen), and we provide to the network the mask and the image with missing part. The model will be trained in this way. We will consider small model, typically a U-Net, and small dataset, such as CIFAR-10 or CIFAR-100, Tiny ImageNet, Oxford-IIIT Pet. A comparison with a model trained without these masks will be conducted.

Supervisor: mailto:Vannary.Meas-Yedid-Hardi@pasteur.fr
Number of Students: 2

Accurate measurement and categorization of tissue characteristics are crucial components of renal tissue histopathological evaluation. In routine clinical assessments, for instance, biopsy quality is assessed by counting glomeruli, and kidney transplant biopsies are thoroughly evaluated using the Banff classification system. Similarly, chronic damage is typically evaluated by visually estimating the extent of interstitial fibrosis and the proportion of atrophied tubules. However, the semi-quantitative assessment is subject to large variability and may lack precision thus may impact the patient treatment. With the advent of deep learning, significant advancements have been made in the field of medical image segmentation, leveraging the powerful capabilities of neural networks.

The aim of this project is to develop a method based on deep learning approach to better stratify kidney disease severity. We will leverage the good performance of machine learning algorithms to quantify and characterize the extent of histologic damage associated with renal diseases. The work will consist in finding the best model to segment the glomeruli. A comparison of well-known models [1,2] and more recent one [3] will be performed, and a special attention will be given to the evaluation metrics.

[1] Ronneberger, O., Fischer, P. & Brox, T. U-Net: convolutional networks for biomedical image segmentation. in Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015 Vol. 9351, 234–241 (Springer, Cham, Switzerland, 2015).
[2] K. He, G. Gkioxari, P. Dollar, and R. Girshick, “Mask R-CNN,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 42, no. 2, pp. 386–397, 2020, https://doi.org/10.1109/TPAMI.2018.2844175.
[3] Graham, S., Vu, Q.D., Jahanifar, M., Raza, S.E.A., Minhas, F., Snead, D., Rajpoot, N.: One model is all you need: Multi-task learning enables simultaneous histology image segmentation and classification. Medical Image Analysis 83, 102685 (2023)

Supervisor: mailto:Vannary.Meas-Yedid-Hardi@pasteur.fr
Number of Students: 2

Vision Transformers (ViT) have emerged as a competitive alternative to Convolutional Neural Networks (CNN) that are currently state-of-the-art in different image recognition computer vision tasks. Transformers are a deep learning architecture, based the mechanism of attention and more specifically of the self-attention module, that were originally proposed for sequence-to-sequence tasks of NLP. ViT have recently achieved highly competitive performance in benchmarks for several computer vision applications, such as image classification, object detection, and semantic image segmentation.

This project will focus on the Swin Transformer [1] (https://github.com/microsoft/Swin-Transformer) stands out as the backbone of computer vision, providing unparalleled flexibility and scalability to meet the demands of modern deep-learning models. The goal of this work is to understand the full potential of this transformer by highlighting the key points of the model. The work will address the performance of the Swin models, on public datasets.

[1] Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., & Guo, B. (2021). Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. ArXiv. /abs/2103.14030
[2] Liu, Z et al. “Swin Transformer V2: Scaling Up Capacity and Resolution”. In: International Conference on Computer Vision and Pattern Recognition (CVPR). 2022.

Supervisor: mailto:Anastase.Charantonis@inria.fr
Number of students: 2

Stochastic interpolants are an evolution of diffusion, a training process that generates outputs from noise. Often, when conditioning such diffusion models to produce outputs according to a specific criterion, this information is added to the process implicitly. Stochastic interpolants enable a gradual transformation from one distribution to another.

For this project, students will read, understand, and code stochastic interpolants (https://arxiv.org/abs/2303.08797) and train them to transform the digit 3 from the MNIST database into the digit 8.

Supervisor: mailto:Isabelle.Bloch@lip6.fr
Number of students: 2

Image segmentation can benefit from a preliminary step where a partition of the image into "super-pixels" is performed. Super-pixels are homogeneous regions, that respect the contours of the objects. This project aims to implement a specific method to compute super-pixels, called "waterpixels", that exploits the watersheds (a tool from mathematical morphology) controlled by markers. The method will be tested and evaluated on various images, and compared to other existing methods, such as SLIC.

Reference:
V. Machairas et al., Waterpixels, IEEE Transactions on Image Processing, vol. 24, no. 11, pp 3707-3716, 2015.

Supervisor: mailto:Dominique.Bereziat@lip6.fr
Number of students: 2

The oceans play a significant role in the regulation of the climate. A detailed understanding of its physical mechanisms is of crucial in understanding its evolution. We are interested in the circulation of the upper layer of the ocean, i.e. the surface currents. These depend, in a first-order approximation, on the upper layer’s sea surface height (SSH).

In this project, we want to retrieve the surface currents from the observed SSH. This can be done in a supervised way by training a neural network from ocean simulation data. Given that the express first-order formation corresponds to a spatial derivative, we expect a convolutional Res-Net architecture to be adapted to this task. As a second step, we want to inject physical information into the network, a constraint called "geostrophy", and check if its inclusion improves the performance of the network.