- A deep learning tool was able to classify subtypes of lung cancer with the same level of accuracy as human pathologists, according to a study conducted by Dartmouth’s Norris Cotton Cancer Center and published in Scientific Reports.
Lung carcinoma is the leading cause of cancer death among men and women in the US, the researchers noted. Lung adenocarcinoma is the most common histological type of lung cancer, and this classification accounts for about half of all cases in the US.
Treatment for lung adenocarcinoma is based on the grade and stage of the tumor, both of which are determined by pathologists. However, identifying the histological subtypes in adenocarcinoma can be extremely challenging, as well as subjective.
“About 80 percent of adenocarcinoma cases contain a mixed spectrum of multiple histologic patterns, and the qualitative criteria used for classification tends to induce inter-observer variability among pathologists,” the researchers said.
“Recent advances in artificial intelligence, particularly in the field of deep learning, have produced a set of image analysis techniques that automatically extract relevant features using a data-driven approach.”
The team developed a deep learning model that would automatically classify different types of lung adenocarcinoma on histopathology slides. The group then compared the model’s classification of 143 whole-slide images to the work of three human pathologists.
The results showed that the deep learning model and all three pathologists’ performances were within 95 percent confidence intervals of agreement on every predominant histological pattern.
The team also created a visualization of the histological patterns identified by the model on the whole-slide images. Each of the three pathologists conducted a subjective, qualitative analysis of the histological patterns, and they each confirmed that the patterns detected on the slides were accurate.
These results demonstrate the model’s potential to improve lung adenocarcinoma classification.
“The visualization of our results and a qualitative investigation by our pathologist annotators confirms that our model’s classifications are generally on target,” the researchers said.
“Our model can potentially be used to aid pathologists in classification of these histologic patterns and ultimately contribute to more accurate grading of lung adenocarcinoma.”
The group said that the model could be used in multiple clinical settings, including existing laboratory information management systems, where it could automatically pre-populate diagnoses for histological patterns on slides.
The tool could also accelerate the tumor diagnosis process by automatically requesting genetic testing for certain patients, based on detected histological patterns. This could enable pathologists to diagnose and treat patients faster.
The findings from this study mirror results from other recent research efforts. In December 2018, a team of researchers developed a deep learning model that was able to replicate diagnostic sleep scores for sleep apnea, sleep staging, and limb movements with the same accuracy as human clinicians.
In a separate study from Mount Sinai Icahn School of Medicine, a deep learning tool that detected neurological diseases in CT scans processed the images 150 times faster than human radiologists.
The Dartmouth lung cancer research group noted that their study had some limitations, including that it was conducted on data from a single medical center, so the data may not be representative of all lung adenocarcinoma histology patterns.
The team also acknowledged that their study used a relatively small dataset in comparison to classical deep learning sets, many of which have more than a million unique images in total.
Still, the researchers are confident that their model could accelerate lung cancer classification, diagnosis, and treatment.
“Our study demonstrates that machine learning can achieve high performance on a challenging image classification task and has the potential to be an asset to lung cancer management,” said Saeed Hassanpour, PhD, a lead author of the study.
“Clinical implementation of our system would be able to assist pathologists for accurate classification of lung cancer subtypes, which is critical for prognosis and treatment.”
The group has made their code publicly available to other researchers to support new research and collaborations in this area. In addition, the team has plans to apply the model to other histopathology image analysis tasks for breast, esophageal, and colorectal cancers.
“If validated through clinical trials, our neural network model can potentially be implemented in clinical practice to assist pathologists,” Hassanpour concluded.