Transforming Stroke Care: Mayo Clinic’s AI-driven Approach to Etiology Classification

Our team participated in a science-research project that involved joining a Kaggle competition organized by Mayo Clinic.

Details about the organizator

Name: Mayo Clinic
Line of business: Healthcare
Founding year: 1864
Size: 63 134 employees
Country: United States

Outline of the problem

The challenge was to accurately classify the origins of blood clots in ischemic stroke using histopathological images. The task involved differentiating between two major acute ischemic stroke etiology subtypes: cardiac and large artery atherosclerosis.

The goal was to provide physicians with a reliable tool to assist in stroke etiology classification and prescribe appropriate treatments.

Proposed solution

The proposed solution involved developing a deep learning-based model that accurately classified the etiology of ischemic strokes based on histopathological images.

Applied technologies:

Python – used for creating scripts and organizing the experiment structure.
Pytorch employed for building and training neural networks for classification task.
HistomicsTK – for color normalization techniques,
MONAI -used for intelligent image slicing and preprocessing.

Enhancing Clot Origin Identification through Deep Learning and Mechanical Thrombectomy Analysis.

Before the project, stroke etiology classification was primarily carried out by highly specialized doctors. Mechanical thrombectomy, a standard treatment for acute ischemic stroke, allowed for the analysis of retrieved clots. However, applying deep learning-based methods to predict stroke etiology and clot origin presented challenges due to the unique data formats and large image sizes.

The objective is to enable healthcare providers to better identify the origins of blood clots in deadly strokes, leading to improved post-stroke therapeutic management and reduced chances of recurrent strokes.

The people and tech behind our project.

The team’s success was built on their expertise in different technical areas. First, they were skilled in image processing, which helped them handle and normalize histopathological images despite variations in lighting, quality and staining caused by different sources of data. This ensured consistency and accuracy during analysis.

Next, their knowledge of deep learning and neural networks was crucial in constructing and training highly accurate classification models. They used these models to predict the origins of blood clots in stroke cases based on the digital histopathological images.

To ensure the reliability of their models, the team also had strong skills in model training and testing. They used techniques like cross-validation to evaluate how well their models performed on new and unseen data.

Moreover, the team’s decision-making abilities in regard to deep learning models played a vital role in their success. They knew how to select the right hyperparameters, optimization algorithms, and loss functions, which significantly influenced the model’s performance and the overall results achieved.

The team divided themselves into subgroups based on their tasks and interests. One subgroup took charge of image preprocessing, ensuring the images were prepared and suitable for analysis. Another subgroup focused on building and training the deep learning models, trying out different configurations to improve accuracy. The third subgroup was responsible for submitting their predictions and ensuring that the model’s performance met the competition’s goals and that the model can handle unseen data.

“It was our first encounter with such large images. Some of these images were massive, reaching up to 2 GB in size. It was quite a challenge even to open them, let alone process, analyze, and evaluate them to determine the type of stroke. Handling these images required innovative approaches and techniques to ensure accurate classification.” –

DAC.digital’s team point of view

Unveiling the Project Step by Step.

Expert Insights: Consulting with a Specialized Professor

We spoke with a professor who specialized in a similar field of research for our consultations. In order to better comprehend the difficulty of the work, we sought her expertise and showed her the histopathological images.

Visualizing Complexity: Presenting Histopathological Images

We used online resources to further our understanding by reading medical literature. For instance, we discovered that sometimes certain characteristics, such as for example higher percentage of white blood cells, may reveal the stroke’s origin. Although we gleaned significant knowledge from these sources, we mostly concentrated on creating computer models.

Model Development: Focused on Computer-Based Solutions

As we went along, we focused on selecting the best neural network to handle the photos and developing effective techniques for processing huge (2GB and more) images that couldn’t be input straight into algorithms. For the purpose of removing one-color backgrounds and slicing remaining tissue, we used intelligent methods suited and optimized for the available data.

Advancing Analysis: Refining Picture Classification

As we learned more, we shifted our attention to optimizing the picture analysis and classification algorithms.

Integration Phase: Testing and Component Alignment

The following phase entailed testing and integrating numerous components. The group worked on creating programs to efficiently process the big pathological images. To make the photographs easier to handle for additional investigation, they developed a script to slice the images. The neural network was trained using a different script using the preprocessed, lower resolution photos. Finally, they created an inferencing script that enabled the trained model to make predictions on photos that it had not seen before.

Processed Data Repository: Storage for Experimentation

The team kept these processed images in a folder and used them both for training and validating purposes during the experiments. We trained the neural network, iteratively analyzed the photos, and then used the inferencing script to test the model’s efficacy on Kaggle’s servers.

Results

The team’s efforts resulted in a deep learning model capable of accurately classifying the origins of blood clots in ischemic stroke based on whole slide digital pathology images. Our solution ranked 46th among 896 teams in the Kaggle competition and earned a silver medal. The model’s success indicated its potential to assist healthcare providers in stroke etiology classification.

Key numbers

Ranking: 46th out of 896 teams in the Kaggle competition.
Achieved Medals: Silver medal for accomplishments in the Kaggle competition.

TAs it involved working with images, the key expertise required was in image processing and normalization. The images often came from different institutions, with variations in lighting and staining, making it essential to apply normalization techniques for consistency. Knowledge of image processing methods, normalization, neural networks capabilities and suitability and decision-making techniques was crucial.

Michał Affek Embedded Machine Learning Researcher

Advancing Healthcare Decision-Making through Technology.

In conclusion, our primary aim was to create a tool that would assist healthcare providers in their decision-making process. Our intention was not to replace the expertise of a doctor but rather to complement their skills and knowledge. The developed algorithm could indicate potential stroke etiologies with a high level of confidence, providing valuable insights to aid doctors in making faster and more informed decisions.

If you have an idea for your own healthcare solution, feel free to reach out to us. We believe in the power of technology and collaborative efforts to enhance medical practices and ultimately improve patient outcomes. Together, we can explore innovative solutions that make a positive impact on the field of healthcare.

Estimate your project!

Let’s revolutionize your customer experience together. Get in touch today!