🛠️ Hardware

We started this week with a small shock – the physics lab was closed to prepare for the school term! Having been evicted from the lab, we moved to the benches along the corridor.

This week was about making final adjustments to our mechanisms. To improve upon the previous holder design aimed at fixing the slanted chair in place, we made a mount for the slanted chair holding the stepper motor with the smaller gear. This was surprisingly difficult to produce, as the mount had to fit exactly with the support or the whole chair would wobble. After a few prints, we nailed down the mount.

Instead of a chair, the other stepper motor got upgraded to a throne! It was more stable and fit well with the motor.

And with that, it concludes our hardware development for our Autoscope 😀

💻 Software

We ran tests to gather data on the accuracy and speed of Cellpose-SAM’s counting algorithm, providing a sense of its performance compared to manual counting. However, after a few rounds of counting, we found that the algorithm had significantly underestimated the number of cells in both specimens we tested on. In the image below, there were approximately 42 cells, but the algorithm counted 17!

We then tested a few images from the internet, which Cellpose was able to count perfectly well. This was puzzling and we had trouble figuring out the cause of this discrepancy. Upon closer inspection, the online images had very high contrast and sharpness. This seemed unnatural and was suspected to be done through edits. We tried maximising the contrast and sharpness through a photo editor, and the results were night and day. The algorithm managed to capture 45 cells this time!

Upon more testing, we figured out that the main issue was not the contrast but the resolution. The resolution had to be quite specific for Cellpose to work, where cells must have around 30px diameter for them to be recognised. By lowering resolution for larger plant cells to have smaller pixel diameters, it made smaller cells unrecognisable. On the other hand, higher resolution allowed for accurate counting of small cells as it increases diameter, but made Cellpose ineffective in counting larger cells. Thus, an additional human step is now required in the autoscope process, where the user has to decide which zoom level to start at the beginning. This made the autoscope work much more consistently as we can set a fixed resolution while using human input to determine the cell size as seen in the image.