Machine Learning and biological imaging

Machine Learning plays nowadays a crucial role in computer vision. A significant part of my training and research involves various optimization and Machine Learning techniques. This page presents the projects where I had the opportunity to exploit these powerful statistical tools to analyze challenging biological structures.

Segmentation of biological structures

My work on curvilinear structures segmentation was for instance applied to the delineation of microtubules observed using total internal reflection microscopy (TIRFM), which relies on the excitation of fluorophores in a biological sample by exponentially decaying evanescent waves created by the reflection of an excitation light into the sample.

Many other biological structures, such as neurites, chromosomes, worms... are thin and curvilinear like the microtubules. Their segmentation faces the same challenges: clutter, low signal-to-noise ratio, illumination inhomogeneity, structures at different depth inside the biological sample ...



     [A] TIRFM images [B] voxels clustering [C] detected microtubules

associated publication

Many biology studies rely on the preparation and staining of biological samples, with the aim of revealing and counting biological structures of interest such as cells and cell nuclei. For small biological samples, the structures of interest are counted manually, but this task becomes quickly very time consuming when the number of available samples increases. Machine Learning based segmentation methods were introduced to address this issue, but the mistakes made by the machine can sometimes significantly deteriorate the results, in particular for needle in a haystack type of problems, which require to detect and count few key structures per image.

We proposed to tackle needle in a haystack segmentation problems by developing semi-automatic software learning to discard obvious uninteresting structures from the biological images. This approach allows the biologists to focus on the most interesting part of their images and prevent/correct the few mistakes generated by the software.

We applied this idea by developing a software, Nuquantus, dedicated to a specific case of cell counting, where the cells of interest are marked by a specific combination of non-specific fluorophores staining various structures of lesser interest.

Nuquantus pipeline for semi-automatic cell counting.

associated publication

Internship Machine Learning project: Fast Shared Boosting

We significantly improved the computational speed of a multi-class variant of Boosting known as shared Boosting by selecting the weak classifiers combined by the Boosting according to Otsu's method, and by partially randomizing their choice. The resulting algorithm can be considered as a hybrid method, bearing similarities with standard Boosting and Random Forests.

associated publications