The competition aimed to gather solutions for current limitations in biomedical imaging related to cell segmentation using deep learning models.
Cell segmentation is usually the first step for single-cell analysis in microscopy in biology and biomedical research. However, it is also commonly considered its most difficult task. That’s why one of this year’s NeurIPS’s competitions focused on Cell Segmentation and deep learning. Its goal is to address weakly supervised task settings, like limited labeled patchets in the sample or many unlabeled images, which are very common in practice.
Eric Upschulte from Forschungszentrum Jülich (FZJ) received a Winner Finalist Award for his submitted solution `Uncertainty-Aware Contour Proposal Networks for Cell Segmentation in Multi-Modality High-Resolution Microscopy Images`. He presents a simple framework for cell segmentation, based on uncertainty-aware Contour Proposal Networks (CPNs). It is designed to provide high segmentation accuracy while remaining computationally efficient, which makes it an ideal solution for high throughput microscopy applications. Each predicted cell is provided with four uncertainty estimations that give information about the localization accuracy of the detected cell boundaries. Such additional insights are valuable for downstream single-cell analysis in microscopy image-based biology and biomedical research.
In the context of the NeurIPS 22 Cell Segmentation Challenge, the proposed solution is shown to generalize well in a multi-modality setting, while respecting domain-specific requirements such as focusing on specific cell types. Without an ensemble or test-time augmentation the method achieves an F1 score of 0.8986 on the challenge's validation set and a very close 3rd place on the hidden test set; 2nd best runtime on par with the best team.