Lucero takes the first step toward pre-isolation qualification of spheroids using Lucero AI.
We are happy to share our latest results on spheroid detection and classification using our convolutional neural network that was trained with images of spheroids taken on our very own prototype (video below)!
We used a popular machine learning strategy where a model developed for a task is reused as the starting point for a model on a slightly different task. Our start point was a network that was initially trained for object detection and classification tasks with everyday objects such as cats, apples, bicycles, and hot dogs and re-trained it to detect spheroids and debris within our microfluidic environment.
“Normally, neural networks require thousands or even hundreds of thousands of images to learn the task at hand and start giving accurate detections. However, given the small number of images that we had to start detecting these precious spheroids, training a convolutional network from zero would have been a much more challenging task. Developing neural network models from zero requires vast computational powers and time. “Transfer learning” was just the strategy we needed to get accurate detections with a small number of images and with very little time spent on training the network.” – Alejandro Diaz Tormo, Co-founder, Lucero Bio
Lucero is still in the middle of the data gathering process. We will continue to work with our partners at SciLifeLab, Karolinska, and KTH who provide the spheroids used to develop our image database and train our AI.
Looking ahead, Lucero’s next milestone is to increase the detection speed of our network to be able to detect spheroids in real-time from our system’s camera feed.