Huge thanks to everyone involved! This has been a fun project—useful for our own experiments and, hopefully, valuable for others as well!

Samplify can also provide detailed parameters of your seed population—size, color, shape, and more. For further information, check out our preprint.

Naturally, this reduces user bias—here, mostly my own. So we built a full framework for training models and annotating images, making it easy for anyone to create their own seed classifier.

This approach to quantifying the triploid block worked remarkably well, achieving near-perfect classification—closely matching human annotations of seed abortion:

Luckily, I already had thousands of manually scored seed images— over 13,000 seeds in total. This gave us a solid base to train a random forest classifier, which our bioinformatician used to achieve reliable seed classification.

In a great collaboration with our institute’s bioinformatics group, we developed "Samplify"—a tool that blends classic image segmentation with the transformer-based Segment Anything Model (SAM) to accurately count seeds, even when they touch or cluster, aligning almost perfectly with manual counting

We always dreamed of automating seed phenotyping to save time and reduce user bias. But Arabidopsis seeds are tiny and often cluster, making analysis tricky—traditional imaging methods usually just excluded those clusters.

Quantification of these aborted seeds was traditionally done manually, i.e., by taking a picture and counting how many normal, abnormal, and totally aborted seeds were present in a given image. The problem was always the huge variation between different people doing the counting.

Our lab has a keen interest in hybridization barriers, such as those that occur in Arabidopsis when parents of different ploidy are crossed — a phenomenon often referred to as the "triploid block." Normal seeds appear light brown, plump, and healthy, whereas seeds from interploidy crosses do not.