Machine & Deep Learning

Biomedical Imaging


Detection, segmentation and classification are usual tasks in biomedical image analysis. The objects of interest (to detect, segment and/or classify) are often heterogeneous in appearance and therefore difficult to describe using standard image features, which has made deep learning a leading approach to providing effective solutions in this area. However, deep learning algorithms bring their own requirements and challenges with them, such as the need for large amounts of annotated images for model training and validation. Such annotations have to be provided by biomedical experts, but this task is time-consuming, error-prone and expert-dependent, especially if segmentation is required.

Our research aims to develop approaches that are robust to imperfect and incomplete annotations in order to reduce the need for the number and quality of annotations, as well as relevant methods for evaluating models in this context (granted by Walloon Region, FEDER and ULB). 

Our fields of application concern :

  • Whole slide imaging and characterization of tissue-based biomarkers in digital pathology
    Whole slide imaging
  • Weak supervised in-vitro phase contrast cell tracking : The project aims to provide an object tracking system based on 2D sequences of phase contrast image of cell. The originality of this project is to rely on a very small supervised set of images, and a diverse data augmentation.
    Weak supervised in-vitro phase contrast cell tracking
​​​​

Other fields

Document analysis

Research in the area of document analysis at LISA addresses questions concerning the analysis of complex schemes and symbols for recognition and similarity evaluation tasks, involving both structure and semantic analysis from 2D documents. 

Applications related to the document analysis:

  • Piping & Instrumentation diagram (P&ID) analysis; P&ID is a an articulate drawing of a processing plan, including basic symbols (e.g. alphanumeric characters) and domain-specific symbols (e.g. related to railways, industrial piping or bio-pharmaceutical production plant). Geometrical Deep learning is used for complex graph comparison. This research is supported by Innoviris (in collaboration with the industry).
  • Detection of trademark infringement using evaluation of intellectual property relevant similarities between logos; Deep learning is used both on the image side for image/image comparison but also for extracting semantic content from text of law associated with court decisions. The project is supported by an ARC project in collaboration with the Juris Lab of the ULB.
  • Digitization and quality improvement of old seismic signal records on paper; the aim is to make the resulting historical data FAIR (Findability, Accessibility, Interoperability, and Reusable). Deep learning is used for image filtering and ambiguity resolution (line crossing, stains etc), this research is supported by Belspo and is done in collaboration with the Royal Observatory of Belgium.

Image super-resolution and image fusion

As the image sensors are becoming cheaper with an increasing resolution, some wavelengths remain more restricted, in particular thermal sensors that remain expensive with respect to their resolution. In that context, we developed an image superresolution algorithm that combines several imaging modalities leveraging the versatility of the deep neural networks and their ability to generale rich and coherent content. Another “beyond the visible spectrum” application consists of generating a RGB visible image from a thermal image alone, extending the night vision possibility.
In collaboration with the VUB (ETRO), we extend this approach to 2D images generated by arrays of microphones; these acoustics images can also have their resolution computationally augmented using these deep neural networks architecture.

Example of network architecture used for super-resolution
Example of network architecture used for super-resolution
Example of super-resolution result obtained on a RGB image (compared with simple bicubic interpolation)
Example of super-resolution result
Thermal image colorization (thermal / night rgb / synthetic rgb daylight)
Thermal image colorization - Thermal Thermal image colorization - Night RGB Thermal image colorization - Synthetic RGB Daylight



Some publications

Other

Biomedical Imaging


Cell tracking: Specific image acquisition and analysis software packages for phase-contrast microscopy were developed for the analysis of cell behavior, in terms of motility, division and death.  The module ivctrack is available on github.



Image registration: Our research aims to develop novel techniques for efficient image registration either at high resolution level and for very large images, or between different imaging modalities with very different resolution levels (e.g., macroscopic vs. microscopic, in vivo vs. histology). This ability is of utmost importance in various medical applications, such as analyses of histopathological biomarker colocalization and therapy effects at the tissue and cell level.
Image registration

Modeling : Developing and validating a mathematical tumor growth model driven by patient-specific Imaging data to improve surgery and radiotherapy planning
 

Other fields

Industrial vision:
  • Various image analysis projects have been carried out in collaboration with our industrial partner Macq-electonics on traffic analysis (vehicle detection, speed estimation, …).
Patient care
  • Arturo project (archive)
    People who undergo amputation develop severe pain they locate in the amputated limb: it is called phantom pain. As a therapy, patients can use a mirror to create the illusion of seeing the amputated limb when in reality they are looking at the reflection of the healthy limb. To make the illusion more realistic, we developed an augmented reality system based on images provided by an RGB+d camera. This system constructs a textured mesh to apply a mirror effect and display a realistic 3D image of the patient, in stereovision on a 3D TV or in an immersion helmet, while tracking in real time the central position of the patient's body
  • Massai Project (archive)
    Based on 3D sensors (time of flight, structured light, stereo), we developed a system for the automatic fall detection of elderly people, people in revalidation and people with psychogeriatric disorders in hospitals, rest homes , rest and care homes and service flats. The project was carried out in collaboration with Mintt company.


Some publications
Updated on April 14, 2022