Source: kalhh/Pixabay
One of the most promising uses of artificial intelligence (AI) machine learning is digital diagnostics for the medical and health care fields. A new study published in the science journal Cell Reports Medicine demonstrates how AI can predict lung cancer from digitized patient tissue samples rapidly and with a high degree of accuracy.
“AI-based approaches to image analysis might be a foundation for useful diagnostic, prognostic, and predictive tools in pathology and oncology,” wrote University of Cologne’s Faculty of Medicine and University Hospital Cologne researcher and corresponding author Dr. Yuri Tolkach along with professor Dr. Reinhard Büttner and their team of research colleagues consisting of Carina Kludt, Yuan Wang, Waleed Ahmad, Andrey Bychkov, Junya Fukuoka, Nadine Gaisa, Mark Kühnel, Danny Jonig, Alexey Pryalukhin, Fabian Mairinger, Franziska Klein, Anne Maria Schultheis, Alexander Seper, Wolfgang Hulla, Johannes Brägelmann, Sebastian Michels, Sebastian Klein, and Alexander Quaas.
Lung cancer is the leading cause of cancer deaths and caused roughly 1.8 million deaths globally in 2022, according to a joint report by the International Agency for Research on Cancer (IARC) and the American Cancer Society (ACS). In 2024, the American Cancer Society estimates that lung cancer will cause 125,000 deaths and there will be more than 234,000 new cases in the United States alone.
Cancer is a disease that happens when abnormal cells divide uncontrollably, and there are over 100 types of cancer, according to the National Cancer Institute. Cancers are named from the part of the body where it originated. Lung cancer is cancer that started in the lungs.
There are two types, non-small cell lung cancer (NSCLC) that makes up 80% to 85% of all lung cancers, and small cell lung cancer (SCLC) that accounts for 10% to 15% of all lung cancers, per the American Cancer Society (ACS). Non-small cell lung cancers include adenocarcinoma, squamous cell carcinoma, large cell (undifferentiated) carcinoma, large cell neuroendocrine carcinoma (LCNEC), adenosquamous carcinoma, and sarcomatoid carcinoma per the ACS. For this study, the researchers focused on non-small cell lung cancers, the most common type of lung cancer.
“We train the primary multi-class tissue segmentation algorithm on a substantial, high-quality, manually annotated dataset of whole-slide images with lung adenocarcinoma and squamous cell carcinomas,” the scientists wrote.
In artificial intelligence machine learning, the quality of the training data impacts the AI algorithm’s overall performance. The researchers trained their main segmentation AI algorithm using a large, labeled dataset. Their AI algorithm was developed using whole-slide image (WSI) data from The Cancer Genome Atlas (TCGA) lung adenocarcinoma and lung squamous cell carcinoma cohorts. The Cancer Genome Atlas is a collaboration between the National Institutes of Health (NIH) National Cancer Institute (NCI) and the National Human Genome Research Institute that started in 2006 and has resulted in over 2.5 petabytes of data describing 33 different tumor types, including 10 rare forms of cancer, based on paired tumor and normal tissue sets from 11,000 patients collected over 12 years. According to the National Cancer Institute, the Cancer Genome Atlas has helped to identify the molecular basis and genomic underpinnings of cancer, tumor subtypes that impact how cancer is classified, and genomic characteristics of tumors that can serve as therapeutic targets for drug development.
To evaluate and validate their AI computational pathology platform for analyzing hematoxylin and eosin (H&E)-stained tissue sections for non-small cell lung cancer, the team used data from a large, international independent multi-institutional cohort.
According to the researchers, their AI algorithm outperforms other studies when it comes to the construction precision of segmentation maps and reported a Dice score for epithelial-only tumor segmentation of 88.5%.
The Dice score, also called the Dice Similarity Coefficient or Sørensen-Dice similarity coefficient for image segmentation, measures the similarity between two datasets. In the field of AI, Dice scores help evaluate the similarity between predicted and actual segmentation.
According to the researchers, they also achieved “the first AI-based algorithm for necrosis density quantification in lung cancer and show its independent prognostic value.”
The scientists point out that earlier studies have established the value of tertiary lymphoid structures in predicting lung cancer and other cancers. Tertiary lymphoid structures are immune cell clusters located in areas that have persistent inflammatory stimulation such as tumors. Their AI algorithm is capable of quantifying tertiary lymphoid structures (TLSs) with a high Dice score of 93.7%.
“The developed computational platform for NSLSC allows for highly precise, quantitative, and objective analysis of tumor morphology,” the researchers concluded.
Copyright © 2024 Cami Rosso. All rights reserved.
