If recent research is any indication, artificial intelligence (AI) has a bright future in medicine. Nvidia developed an AI system that can generate synthetic scans of brain cancer. Google subsidiary DeepMind has demonstrated a machine learning algorithm that can recommend treatment for more than 50 eye diseases with 94 percent accuracy. And in newly published research, New York University (NYU) showed how AI might aid in lung cancer diagnosis.
A paper today published in the journal Nature Medicine (“Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning”) describes how a team of NYU researchers retrained Google’s Inception v3, an open source convolutional neural network architected for object identification, to detect certain forms of lung cancers with 97 percent accuracy.
That’s encouraging news for the more than 200,000 people diagnosed with lung cancer each year. More than 150,000 people die annually as a result of disease-related complications, according to the American Cancer Society and the Cancer Statistics Center.
“[T]his computational approach could play a role in both routine tasks and difficult cases … in order to allow the pathologist to concentrate on higher-level decisions, such as integrating histologic, molecular, and clinical information in order to guide treatment decisions for individual patients,” the team wrote.
They trained, validated, and tested Inception v3 on 1,634 whole-slide images from The Cancer Genome Atlas (TCGA), a public dataset maintained by the National Cancer Institute (NCI) and National Human Genome Research Institute (NHGRI) that maps genomic changes in 33 types of cancer. Once it began reliably — with 99 percent accuracy — identifying cancerous cells, the team taught it to tell the difference between two of the most common forms of lung cancer: adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC).
It performed well when let loose on independent samples taken from cancer patients at NYU, correctly diagnosing images between 83 and 97 percent of the time. That’s despite encountering features it hadn’t previously seen, including blood clots, blood vessels, inflammation, necrotic regions, and regions of collapsed lung.
All the more impressively, the model took just 20 seconds on average, running on a single-GPU PC, to calculate classification probabilities.
Next, the researchers trained it to identify not just cancerous tissue, but the genetic mutations within the tissue. Fed images from TCGA and genetic profiles for each tumor, the system drew on subtle differences in the tumors’ appearance to predict six of the ten most commonly mutated genes in LUAD (STK11, EGFR, FAT1, SETBP1, KRAS, and TP53).
“These cancer-driving mutations appear to have microscopic effects that the algorithm can detect,” Aristotelis Tsirigos, a pathologist at the NYU School of Medicine and a lead author on the new study, told Wired. “What those subtle changes are, however, we don’t know. They’re buried [in the algorithm] and nobody really knows how to extract them.”
In the future, the team hopes to extend the model’s classification to other, less common lung cancers such as large-cell carcinoma, in addition to “histological subtypes” and features such as necrosis, fibrosis, and others. According to Wired, they’re considering forming a company and seeking Federal Food and Drug Administration (FDA) approval to commercialize their technology.
“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 wrote. “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 that aims at developing a multiplex approach with patient-tailored therapies.”