AI for Live Mouse Cancer Progression Detection

Overview

NIST has partnered with the National Cancer Institute (NCI) at the National Institutes of Health (NIH) to study cellular characteristics and cellular movements in live mice that indicate the presence of tumors. NIST researchers have developed neural networks designed to automatically detect, separate, and track cells in live mice in 3D over time. This AI-driven imaging approach can dramatically accelerate oncology research and lower research bottlenecks, paving the way for faster, more targeted cancer treatments.

Background

The National Cancer Institute (NCI) at the National Institutes of Health (NIH) uses modern microscopy to detect tumor formation and progression in living organisms. This approach generates millions of images acquired at a wide range of scales and from subcellular to whole organs. NCI requires innovative image analysis methods to identify and measure features across millions of images and to relate these measurements to their disease state.

Analysis of the NCI datasets has, until recently, depended on computer vision experts (specialists trained in getting computers to identify images) to manually specify how to detect cellular and subcellular features, such as mitochondria, organelles, and nuclei. Over the last decade, these methods have given way to deep learning techniques that depend on neural network models. The models are first trained to first detect simple lines and edges, and then to recognize more complex features. Because of increased computing power and freely available machine learning frameworks like TensorFlow and PyTorch, these models can now be trained relatively quickly- leading to an explosion of research. 

Approach

The project team is combining emerging AI techniques with traditional feature engineering and tracking in cell biology applications, to generate image measurements that help characterize tumor formation and progression. The team is applying an innovative unsupervised training approach that lets a neural network use image features extracted from a full 3D cell model to learn efficiently with limited recourse to ground truth measurements. 

The research team is engaged in the following activities:

1. Create training sets by iterative learning for a 3D single cell neural network model to detect single cells in the entire tongue. Extract features on a single cell level and compare these features between cells in normal tissue and cells in tumor tissues.
2. Continue to research automated feature engineering in neural networks for distinguishing tumor cells from normal cells and relating them to biologically meaningful 3D features in order to improve interpretability of AI models.
3. Understand and optimize a wide range of hyper-parameters that impact accuracy and execution speed of multiple tasks as analyses expand from 2D to 3D, 4D, and 5D, such as class weighting schemes, image normalization schemes, use of multi-channel stains, image tiling, and data augmentation in 3D.

Impact

By developing AI models for segmenting 3D images of mouse tongues and classifying regions as cancerous tumors, this work advances the intersection of AI-enabled science and biomedical imaging at NIST.