If you specialize in pancreatic cancer, you know that it’s notoriously difficult to diagnose—clinical symptoms are easy to overlook, and they may not show up until the late stages of the disease. NCI-funded researchers at the Mayo Clinic have created a tool that could make early diagnosis easier in the future, and it centers on an automated Convolutional Neural Network.

The researchers’ tool has some key advantages over earlier models. The model:

• stands up to testing on a very large data set (i.e., using 1,776 Computed Tomography [CT] scans from patients with diverse demographics and images obtained with a variety of machine settings).
• identifies undetectable cancers on CT scans of seemingly normal pancreases up to 438 days before clinical diagnosis.
• shows similar results on both large and small tumors.
• offers a fully automated approach that requires little-to-no human intervention.

Corresponding and senior author, Dr. Ajit H. Goenka, said, “We know CTs harbor ‘visually occult’ cancers; that is, cancers that are present on the CT, but which are beyond the scope of human perception. By training our model on such a large and diverse amount of CT data, we’re able to distinguish changes in pancreatic texture that could help in identifying cancer months earlier than our current standard-of-care methods.”

Such a model could be a key part of diagnostics in the future. Said Dr. Goenka, “Our model could serve as a second ‘set of eyes,’ helping radiologists and other specialists see subtle or small cellular changes that signal the presence or potential for pancreatic cancer.”