Machine Learning

Our team has industry experience dating back to the 90s - from artificial life to computer vision, we've got you covered.

TensorFlow Keras OpenCV Dagster AWS YOLO Python GitLab
TensorFlow Keras OpenCV Dagster AWS YOLO Python GitLab

Artificial intelligence is the new hotness 🔥

Read on for some of our recent work...

Natural Language Processing

Large-scale PHI de-identification

In a recent project, we faced the challenge of de-identifying large amounts of private health information while preserving as much contextual information as possible.

Traditional redaction techniques would destroy important data, making it unusable for our machine learning models. To address this issue, we employed a technique called de-identification, also known as jittering.

De-identification involves replacing sensitive information such as names and dates with fake information, while maintaining contextual information such as gender, ethnicity, and time between appointments.

We were able to effectively preserve the contextual information in our data while still protecting patient privacy. This allowed us to use the data to train our client's ML models.

Computer vision

Wound surface area calculation

Acute and chronic wounds are an economic burden to healthcare systems around the world. Wound after-care and follow up appointments take up doctors and patients valuable time.

We comprised the ML team on a recent healthcare project, developing a modular AI system capable of automatically calculating the surface area of wounds using machine learning and computer vision.

Along the way we combined "old-school" computer vision solutions with cutting edge convolution neural networks for image segmentation. This two pronged approach proved very successful, allowing the system to accurately track wound healing.

Schematic diagram: Pipeline for the preparation of a wound image dataset for a machine learning model within the healthcare industry.

Real-time object recognition

As an internal research project we built an object recognition system using the freely available CIFAR-100 dataset.

The model took over 22 hours to train on a high-end consumer graphics card.

In the end we achieved 74% accuracy and the ability to identify 100 categories of objects. At the time this was competitive with other cutting-edge techniques.

The training and refining of the machine learning model proved time consuming and computationally expensive. However, once optimised the model was suitable for real-time use on a smartphone.

Get in touch

Call us on +44 (0)117 325 1055