Could AI Automate Diagnosis (& Targeted Tx) of Lung Cancer?

Kristin Jenkins

September 19, 2018

An artificial intelligence (AI) tool that can diagnose non–small cell lung cancer types from histopathology slides with 97% accuracy and can predict the mutational status of six driver oncogenes in seconds could automate the diagnosis of common lung cancer types and increase the use of targeted therapy, according to researchers.

"Our approach can be applied to any cancer type," say Aristotelis Tsirigos, PhD, Department of Pathology, New York University School of Medicine, New York City, and colleagues in a report published online September 17 in Nature Medicine.

The large, convolutional neural network model, which was developed by Google, is capable of carrying out complex visual tasks using microarchitectural units that process data in multiple resolutions. Called "inception," this network has already been taught to classify skin cancers and to detect diabetic retinopathy, as previously reported by Medscape Medical News.

For the lung cancer study, the investigators trained inception to distinguish tumor from normal lung tissue. They then developed models to classify 1634 whole-slide images from the Cancer Genome Atlas (TCGA) as either adenocarcinoma (LUAD), squamous cell carcinoma (LUSC) or normal lung tissue.

After validating their results in independent cohorts, the investigators pitted the diagnostic power of their neural network models against the combined clinical acumen of two thoracic pathologists and one anatomic pathologist. All three were asked to independently classify the whole-slide images by visual inspection alone.

"The performance of our method is comparable to that of pathologists," the authors say. Typically, lung biopsy is used to diagnose lung cancer type and stage, they note.

The study showed that the model distinguished normal tissues from tumor tissues with an average area under the curve (AUC) of ~0.99, demonstrating a sensitivity and specificity of 0.97 AUC. This is a significant improvement over the diagnostic accuracy of previous deep-learning models, the researchers point out. They note that earlier reports demonstrated an AUC of 0.75 and 0.83 in the classification of tumor slides from the TCGA.

Surprisingly, AI was also able to identity some mutations that are treated with trageted therapy.

When the neural network was trained to use whole-slide pathology images to predict the 10 most commonly mutated genes in LUAD, it was able to accurately predict six: STK11, EGFR, FAT1, SETBP1, KRAS, and TP53.

EGFR mutations are present in about 20% of LUAD, while anaplastic lymphoma receptor tyrosine kinase (ALK) rearrangements are present in <5%, the researchers point out. Although mutations in the KRAS gene are found in about 25% of LUAD tumors and TP53 mutations in about 50%, both present challenging drug targets.

Notably, about half of the slides that were incorrectly classified by the model were misclassified by at least one of the pathologists.

However, 83% of the slides that were incorrectly classified by at least one of the pathologists (45 out of 54) were correctly classified by the algorithm, the study showed.

After the researchers measured the agreement between the TCGA classification, the deep-learning model, and the pathologists, both individually and collectively, they found that the algorithm was in agreement with TCGA more often than the pathologists. This finding did not reach statistical significance.

In addition, the algorithm was a lot faster than the pathologists. It was able to process a slide in as little as 5 seconds, whereas the pathologists took up to 5 minutes.

"Overall, this study demonstrates that deep-learning convolutional neural networks could be a very useful tool for assisting pathologists in their classification of whole-slide images of lung tissues," the researchers write. "This information can be crucial in applying the appropriate and tailored targeted therapy to patients with lung cancer, increasing thereby the scope and performance of precision medicine...."

Because treatment guidelines for LUAD and LUSC differ, classification of lung cancer type is a key part of the diagnostic process, they point out.

In the absence of definitive histologic features, the use of immunohistochemical staining to distinguish one cancer type from another can be time consuming and challenging, particularly when the tumor is poorly differentiated. Results from genetic testing can take up to 4 weeks.

The development of new, inexpensive graphics processing units has made it possible to train larger and more complex neural networks to segment or classify medical images, particularly for whole-slide images. Applications currently include breast cancer diagnosis, glioma grading in brain cancer, colon tumor analysis, identification of epithelial tissue in prostate cancer, and diagnosis of osteosarcoma.

The researchers emphasize that although these algorithms may provide helpful prognostic information in the initial diagnosis of lung cancer, the next steps of staging the tumor and predicting treatment response remain the domain of the pathologist.

In an interview with Medscape Medical News, Tsirigos, a computational biologist who specializes in cancer epigenetics, genomics, and image analysis, predicted that pathologists won't be replaced by deep learning any time soon.

"Will you ever want the AI system to make the final decision? I doubt it," he told Medscape Medical News. "It's like a second opinion and more about error avoidance, with the doctor signing off on the case."

During this lung cancer study, pathologists "embraced this approach," he added. "They came to us and asked if they could work with us because it was so interesting and going to help them so much. My message to clinicians is to feel empowered, not threatened."

Tsirigos pointed out that the researchers trained the system with about 1500 slides. If they train the system with 15,000 slides, "it would be more accurate," he said.

The fact that the algorithm could predict the mutational status of a driver oncogene just by looking at the whole-image slide surprised the researchers. "Typically, if you want to screen for mutations in a clinic, results might take 2 to 4 weeks," Tsirigos noted.

Many physicians choose to initiate therapy rather than wait, he said. As a result, potential candidates may not receive targeted therapy or be enrolled in clinical trials. "If you have the data and the algorithm, you get the genetic answer in seconds. It would be like getting an instantaneous alert from the pathology department saying, 'This is a good candidate for targeted therapy.' "

You get the genetic answer in seconds. It would be like getting an instantaneous alert. Dr Aristotelis Tsirigos

Currently, the researchers are exploring the use of multiple data modalities, including metadata, clinical data, and MRI data, to predict the response to therapy.

The code to the convolutional neural network model is available online.

Dr Tsirigos and coauthors have disclosed no relevant financial relationships.

Nat Med. Published online on September 17, 2018. Abstract


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.