Home

Dry Lab Notebook

At the start of the project, we conducted a literature review of multiple object detection models. Through our research and after discussions with professors of the field, we concluded in using the Yolo architecture for our system. In collaboration with the wet lab team, we defined the way images need to be taken, explained how important the annotation of the data is and that we needed a large number of images to train the network the correct way. Furthermore, we get ourselves in implementing a modification of the Yolo architecture since our model is much smaller, has a different number of cells and anchors, and validates the implementation using a single annotated image.

In about June, we had a small set of images from the electrophoresis gel results, and we used them to train our network. Initially, since we had only 24 images, we used them all as a training set to validate the system again. When we split the dataset into training and testing set using 19 and 5 images, respectively, we find out that the network can’t make predictions in images that are not trained in, even though they look similar. This happened because the instances’ distribution was unbalanced, and the pictures were taken from different heights and with varying conditions of light. So we decide to crop the images to contain only the gel area and generate new images using a photo editing program, photoshop, to make some columns have representation with instances.

In July, we decided to generate more images with variations in contrast and brightness. Another augmentation technique that was applied was Morphological transformations, such as Erosion and Dilation. We generate more images with variations in contrast, brightness, and different kernel sizes in erosion and dilation for every dataset image. The results still proved the same, so we decide to use images from different crops, different zoom levels, and feed the network with the original images to have more instances.

In September, the first results were encouraging as we used some new images provided by the wet lab team to test the network. The final idea that came up to us was to use grayscale images instead of RGB to make the background color independent of the classification result.

Finally, in early October, we finish our work in Yolo, and we implemented the AI tool, a web-based application that assists the user through its step-by-step guidance and the dosage recommendations it provides.

Team:Patras/Dry Lab Notebook

Dry Lab Notebook

Sponsors

About

Follow Us