Crystal Morphology Classification

Problem

Crystal habit, or morphology, is a crucial attribute in drug manufacturing, denoting how easily drug crystal formations can be pressed into pill form. Microscopic images of drug crystals are generally evaluated and identified by subject matter experts, a time consuming process which creates a bottleneck in the process.

Solution

Deep learning presents an opportunity to eliminate bottlenecks with partial or complete automation of image classification tasks.  Retina is particularly suited to handle these complicated tasks that are beyond simple binary classification tasks like the classic cat vs. dog.  ImageVec, the preprocessing and algorithm suite behind Retina, has been designed with life science and pharmaceutical applications in mind.

Often a lack of annotated data can be a roadblock to properly leveraging deep learning. The answer to this lies in transfer learning, starting with a large pre-trained, general knowledge model and then fine tuning on just a small amount of user data for downstream tasks. ImageVec leverages a number of state of the art computer vision transfer models as the base of its capabilities. This greatly reduces the amount of user data needed to get up and running.

Images of microscopic slides are notoriously difficult to process for successful modeling, but Retina’s custom built preprocessing pipeline really shines in this case. Each image was tiled into subsections and then those subsections were filtered for complexity, eliminating white space around the samples. Handwritten notes with potential label leakage were scrubbed from the images. Finally, the data was resampled and the minority classes were augmented with a combination of rotation and added noise. After processing the data, models trained with Retina were able to achieve over 90% accuracy identifying multiple different crystal shapes.