Embedded Files
H&E whole slide image of lungs collected from a mouse model of lung adenocarcinoma. Hover to show manual tumor segmentation and grading. The large purple structure on the left of the slide is lymphoid tissue that is not part of the lung.
Mouse models of lung adenocarcinoma
Mouse models are widely used to study lung adenocarcinoma (LUAD), the most common form of lung cancer (especially in non-smokers). Researchers can manipulate the genomes of mice to mimic mutations found in human patients with relative ease. However, most of the approaches used to model LUAD in mice cause numerous primary tumors (i.e., not metastases) throughout the lungs, a feature not common in human patients. The independent development of these tumors allows researchers to get more data from fewer animals. Still, analysis of each tumor is a time-consuming process. There are computer programs that can assist in identifying tumors from a microscopy slide. Yet, many analyses, such as tumor grading (assessing the severity of the tumor), must be done manually. Because the grading of tumors is based on identifications of specific patterns and features, we reasoned that a computer could perform this task with high accuracy and faster than a human.
Overview of training data generation (top) and network architecture of GLASS-AI (bottom). Taken from (Lockhart et al. 2023 npj Prec Onc).
Building a machine learning model for tumor grading
Training a good machine learning model requires good data, usually a lot. We built our training dataset using annotations (i.e., segmentation and grading of tumors) from multiple human raters on whole slide images. We also included three different mouse models of LUAD to increase the generalizability of our software. We designed our machine learning model to recognize six classes in the histology images: Normal Alveoli, Normal Airway, and Grades 1 – 4 LUAD using 36,000 image patches from our dataset. We call this trained network Grading of Lung Adenocarcinoma with Simultaneous Segmentation by Artificial Intelligence (GLASS-AI). We based GLASS-AI on the ResNet18 architecture and wrote the software in MATLAB.
The image/label patch library used to train GLASS-AI has been made publicly available.
You can find the source code for GLASS-AI and links to installers for the standalone version for Windows and Mac on my GitHub page.
GLASS-AI analysis of the whole slide image above with the human annotations overlaid (shown as colored borders). The lymphoid tissue on the left of the slide was excluded from the analysis. Hover over the thumbnail to zoom in.
GLASS-AI reveals intratumor heterogeneity
When human raters annotate a tumor, they assign a single overall grade based on the patterns they observe in the tumor. Typically the overall grade is the highest grade present that makes up ≥10 – 20% of the tumor's area. GLASS-AI, on the other hand, assigns classes to individual pixels of the whole slide image without first identifying tumors, revealing the heterogeneity of tumor grades with unprecedented clarity. Leveraging this higher-resolution grading, we can now perform more detailed investigations into how LUAD tumors progress and respond to new therapies. Using GLASS-AI, we can also accelerate our research by spending less time annotating slides and more time conducting experiments.
This work was published in "Grading of lung adenocarcinomas with simultaneous segmentation by artificial intelligence (GLASS-AI)" (Lockhart et al. 2023 npj Prec Onc).
After publication, we were invited to write up a blog post describing this work for the Nature Cancer Community group. That post (and our paper) were also featured on the Nature Portofolio Instagram page.
Page updated
Google Sites
Report abuse