Transforming stroke care with AI - a Computer Vision solution for Mayo Clinic

Battling stroke with AI innovation

Mayo Clinic is a renowned healthcare organization at the forefront of innovation. This case study explores our team's success in the science-research Kaggle project, which was focused on assisting physicians in stroke care. The concept was to develop a Computer Vision-based solution to etiology classification, in order to quickly identify stroke origin – all organized as a research competition by the Mayo Clinic.

Understanding the challenge: Accurate classification for effective stroke treatment

The project aimed to effectively differentiate between major acute ischemic stroke etiology subtypes - cardiac and large artery atherosclerosis - which is crucial in determining appropriate treatment plans. Mayo Clinic sought to develop a reliable tool that could assist physicians in accurately classifying the origins of blood clots to optimize stroke care.

Leveraging deep learning: Building a cutting-edge model

Leveraging our expertise in AI, Machine Learning, and Computer Vision, we developed a deep learning-based model that could classify the etiology of ischemic strokes from histopathological images.

Model Development: Focused on Computer-Based Solutions

Goal achieved: Optimizing Image Processing 
Our team meticulously selected the most suitable neural network architecture for analyzing histopathological images. To handle large images, intelligent techniques were employed to remove one-colour backgrounds and preprocess the remaining tissue effectively.

Advancing Analysis: Fine-tuning Picture Classification

Goal achieved: Enhancing Accuracy
Continued refinement of the neural network models allowed for improved accuracy in classifying the origins of blood clots in ischemic stroke cases. Iterative adjustments were made to optimize the models' performance and achieve precise classification results.

Integration Phase: Testing and Component Alignment

Goal achieved: Seamless Integration 
A comprehensive integration phase involved testing and aligning various components. To facilitate handling large pathological images, efficient processing programs were developed. These programs included scripts for image slicing, and the neural network was trained using preprocessed, lower-resolution images. An inferencing script was created to enable the trained model to make predictions on new, unseen photos.

Processed Data Repository: Enabling Experimentation

Goal achieved: Accessible Experimentation
To support ongoing experimentation, the processed images were stored in a dedicated repository. These images served as the foundation for training and validation during the development process. The neural network was iteratively trained and evaluated, utilizing the inferencing script to measure its efficacy on new images.

Result: Better stroke care with AI through accurate classification

Punktum's collaboration with Mayo Clinic led to the development of a cutting-edge deep-learning model that accurately classifies the origins of blood clots in ischemic stroke. In the Kaggle competition, the solution achieved an impressive 46th place among 896 teams, earning a silver medal. The developed solution is a great example of how deep tech can pave the way for more targeted and effective treatments, revolutionizing stroke care and improving patient outcomes.

Technologies Used:


Python

Scripting and experiment organization

Pytorch

Neural network development and training

HistomicsTK

This image has an empty alt attribute; its file name is HistomicsTK.svg

Color study and analysis

MONAI

Intelligent image slicing and preprocessing

Or contact us right now and let's build your product together!

Your inquiry