Interpretable Sequential Multiple-Instance Learning for Medical Imaging

Background

Machine learning has quickly gained relevance in the medical field for its ability to analyze complex datasets to increase diagnostic accuracy and is commonly used to analyze individual datasets of images such as x-rays, ultrasounds, and similar medical images. However, traditional deep learning methods typically struggle with the complexity and variety inherent in sequences of these images.

Technology Overview

Researchers at Boston Children’s Hospital have developed an innovative algorithm tailored for interpreting sequential medical images and making predictions for these sequential images, ensuring that each additional slice in a scan meaningfully contributes to the final prediction, especially in determining which slice holds greater importance in the overall diagnosis. This approach significantly improves the accuracy of medical diagnoses while facilitating healthcare professionals' proper diagnosis in the shortest amount of time. Variations in each slice are also integrated to calculate a model uncertainty metric. This metric guides the healthcare professional decision on whether to rely on the model’s prediction, significantly improving the model’s interpretability. 

Applications

• Machine learning analysis of sequential diagnostic images.

Advantages

• Improvements to diagnostic accuracy for medical images
• Increase interpretability of sequential diagnostic images