Predictive AI model might aid early-stage NSCLC planning

An artificial intelligence (AI) algorithm can predict who will develop brain or other metastases from the point of early-stage non-small cell lung cancer (NSCLC), according to research published March 4 in the Journal of Pathology.

The study demonstrated how researchers can take whole-slide images (WSIs) produced from routine hematoxylin and eosin (H&E)-stained NSCLC tumor tissue slides and use them to train an AI model to potentially perform histologic review by pathologists. 

The researchers from Caltech in Pasadena, CA, and Washington University School of Medicine in St. Louis emphasized that their work is important because correctly specifying which patients with stage I NSCLC will not develop metastasis might spare people from harmful and expensive systemic therapy.

"Our study is an indication that AI methods may be able to make meaningful predictions that are specific and sensitive enough to impact patient management," Dr. Richard Cote told Caltech Weekly. Cote is head of the department of pathology and immunology at Washington University School of Medicine and co-principal investigator of the pilot study.

"Overtreatment of cancer patients is a big problem," added Dr. Changhuei Yang, Thomas G. Myers professor of electrical engineering, bioengineering, and medical engineering at Caltech and an investigator with the Heritage Medical Research Institute. "Our pilot study indicates that AI may be very good at telling us in particular which patients are very unlikely to develop brain cancer metastasis." 

Furthermore, based on the regions of interest (ROI) that most strongly contributed to the deep-learning (DL) algorithm’s ability to predict progression versus no progression, it appeared that the AI's basis of prediction relied not only on subtle and complex histologic features of tumor cells but also non-tumor cells and the tumor microenvironment, according to the study's authors.

Brain metastases can occur in nearly half of patients with early and locally advanced NSCLC. However, pathologists lack reliable histopathologic or molecular means to identify those who are likely to develop brain metastases, the authors wrote. 

As a potential AI assistant for pathologists, this DL model started with data and biopsy images of diagnostic WSIs of 158 patients with stages I to III NSCLC (65 metastases positive [Met+], 93 metastases negative [Met-] within five years). All cases were assessed for tumor adequacy. From each scan, regions of high tumor cellularity and surrounding tumor microenvironment were annotated by one reviewing pathologist.

Case outcomes were correlated with DL predictions only after training and validation and subsequent testing processes were complete, according to the authors. Pathologist reviewers were also blinded to outcome and stage data, although they were obviously privy to histologic subtype.

While the study cohort was relatively small, the AI model predicted progression with an accuracy of 87% using separate validation and testing sets, regardless of histologic subtype and stage. An analysis of the stage I cases revealed specificity of 95.7% in predicting Met-, with a negative predictive value of 92.9%. The AI was less accurate in predicting Met+ patients, with a sensitivity of 74.3% but a positive predictive value of 83.3%.

The framework employed in this study allowed investigation of the model’s attention at tile-level resolution over the WSI. "Although the model was trained and tested on 1,000 image tiles ... the prediction scores were aggregated over the entire image to make a slide-level prediction," the authors wrote.

Importantly, the DL model identified features that were not readily discernable by a trained pathologist (such as tumor grade, necrosis, lymphocytic infiltration, spread to airway spaces), and it even outperformed careful histologic review by four expert pathologists, the authors wrote.

The research suggests that DL-based algorithms such as this one could eventually reduce the difficulty of managing early-stage (I, II, and III) NSCLC. Yang cautioned that the work is only a first step and that a larger study is needed to validate the findings. 

Page 1 of 176
Next Page