Conférence d’Anthony Bilodeau, étudiant au doctorat en biophotonique

High throughput quantitative analysis of super-resolution microscopy images is very challenging due to the complexity of the detected nanoscopic structures. When automatic detection methods are not applicable, researchers have to go through the tedious process of manually classifying and segmenting the structure of interest in all acquired images.

We propose a weakly supervised deep learning approach for the automatic segmentation of fluorescence microscopy images. We trained a deep convolutional neural network to recognize and segment simultaneously multiple structures of interest in a super-resolution microscopy image using solely whole image binary labels for training. We then used both low- and high-level learned features to generate a semantic segmentation. Precise segmentation is achieved on a crafted dataset of cluttered handwritten digits and serves as a working proof of concept.

We tested our approach on a more challenging super-resolution microscopy images dataset of the F-actin cytoskeleton. We obtained increased segmentation performances compared to conventional U-Net approaches. Overall, the proposed technique would alleviate the labeling task of semantic segmentation done by researchers as it only requires whole image labels.