We were inspired to explore medical image diagnosis using machine learning and gain practical experience applying machine learning techniques to classify medical data.
What it does
It enables a classification of 2D sagittal image patch containing three verberge and outputs the fracture grade of the middle vertebra.
How we built it
We used Google Colab, Jupyter Notebooks and Python to build a state-of-the-art vertebral body fractures classifier using Keras. The main approaches were a DenseNet121 and a uNet model architecture.
Challenges we ran into
Major challenges that emerged in practice were issues with preprocessing the data, image visualisation and image recognition. By converting images from 16-bit greyscale to 8-bit RGB (by multiplying the greyscale channel), we were able to fix this issue.
Accomplishments that we're proud of
We used iterative experimentation in data processing to find the best model for solving our issues by further improving our classifier model.
What we learned
How to build and evaluate a deep learning classifier model: – hands-on experience using and preparing X-ray dataset for processing. – applying computer vision to extract information from unstructured medical data. – using multiple transfer learning models for image classification. – exploring methods to interpret diagnostic and prognostic models. – visualising learning with GradCAM.
What's next for Diagnosis of Vertebral Body Fractures Group
Most probably another hackathon to take up the next challenge and gain practical experience applying machine learning to concrete problems.