📈 Fuse My Cells Challenge Results¶

In order to advance new methods for 3D image-to-image fusion using deep learning in the fields of biology and microscopy, the Fuse My Cells challenge main task was to predict a fused 3D image using only one 3D view (used for the fusion).

This objective was made possible thanks to the structuring role of France Bio-Imaging, which has allowed the creation of a new database of more than 400 paired 3D images.

For our second challenge, we had 136 registrations and 6 participating teams. After a preliminary test phase - on 3 test images with ~5 submissions to familiarize with the data and the submission procedure- the participating teams evaluated their methods during the (final) evaluation phase on 32 images with a unique submission.

Many thanks to all involved for their hard work !¶

From the data contributors to the challenge participants !

🏆 We are delighted to announce our participants winners :¶

1. lWM with the FuseMyCells algorithm based on two separate RUNet encoder-decoders
   code access

2. CaVa with the Fuse_my_cells_dual algorithm based on a lightweight UNet model with Z-axis attention
   code access

3. Cyril.meyer.68 with the M²N² algorithm based on a multidimensional Gaussian filter
   code access

We chose to use 2 different metrics(normalized in comparison with ground truth) to focus on 2 factors: Pixels for similarity and segmentation for usability. These metrics were calculated on Nucleus and Membranes images. The results obtained were used to establish a ranking for each metric and organelle. For the final ranking, we took the average of the ranks for each metric:

We can see a relatively large standard deviation on average for these ranks. We therefore wanted to take a closer look at the scores.

Here are the average scores and their deviations over the 32 evaluation images for each SSIM and IOU metric.

We can see that, even if pixel-level predictions don't always match the ground truth (the merged reference image), the IOUs are clearly better than the ground truth.

What does this mean? What might it tell us about the relationship between accuracy and generalization in model performance? Join us at ISBI 2025 to explore these implications and delve into the details.