Introduction

Crimson is an application for visually detecting poor quality blood spot samples in newborn screening by utilizing deep learning for accurate sample classification within a robust PWA. Crimson helps patients by enabling timely treatment and reducing parental anxiety. Crimson also offers significant cost savings and freed resources by removing the need for additional sample collection.

Goal

Our goal was to create a proof-of-concept classification algorithm and prototype mobile application for visually detecting poor quality blood spot samples in newborn screening. The algorithm should take into account 5 categories of samples: good, too big, too small, multispotted and compressed. The app needs to be usable also without an online connection.

Results

We developed an image processing and detection pipeline to classify blood spot quality and detect good samples from bad. The pipeline utilizes a collection of image processing technologies, PyTorch depth detection and finally a custom TensorFlow CNN for classifying the samples. For easy access between different devices, the model was implemented on a progressive web application (PWA) and is run locally on the user’s device. A photo can be taken of the sample in addition to adding the image file.