A noninvasive “virtual biopsy” may help diagnose skin cancer without making any cuts.
Stanford Medicine researchers developed a method that uses lasers to penetrate tissue and create a high-resolution, three-dimensional reconstruction of the cells it contains.
From this virtual reconstruction, they can make cross-sectional images that mimic those generated by a standard biopsy.
The new method could be used to noninvasively scan the skin for unhealthy cells as well as provide rapid results on biopsies taken elsewhere in the body.
“We’ve not only created something that can replace the current gold-standard pathology slides for diagnosing many conditions, but we improved the resolution of these scans so much that we start to pick up information that would be extremely hard to see otherwise,” says the study author Adam de la Zerda, PhD, an Associate Professor of Structural Biology at Stanford.
The method was developed by Yonatan Winetraub, PhD, a former graduate student in the de la Zerda lab who now leads his own research lab at Stanford focusing in part on virtual biopsies.
“This has the potential to transform how we diagnose and monitor concerning skin lesions and diseases in the clinic,” adds co-author Kavita Sarin, MD, PhD, an Associate Professor of Dermatology.
For nearly a decade, de la Zerda and his colleagues have been studying optical coherence tomography. Typically used by ophthalmologists to image the back of the eye, OCT scans measure how light waves from a laser bounce off a tissue to create a rendering of its insides.
As de la Zerda and Winetraub enhanced the OCT scans so they would work in organs other than the eye — developing both new hardware to collect data and new processing methods — they needed a way to verify the accuracy of their scans, so they sent the tissues they were scanning with OCT to pathologists to create H&E images.
“We kept improving and improving the quality of the image, letting us see smaller and smaller details of a tissue,” de la Zerda explains. “And we realized the OCT images we were creating were really getting very similar to the H&Es in terms of what they could show.”
The higher resolution of the OCT images opened the door to using the method to diagnose disease without producing H&Es. But de la Zerda and his colleagues thought clinicians would be more apt to use OCT if the images looked familiar.
“Every physician in a hospital is very much used to reading H&Es, and it was important to us that we translate OCT images into something that physicians were already comfortable with —rather than an entirely new type of image,” de la Zerda says.
Winetraub turned to artificial intelligence to help convert OCT scans into flat images resembling H&E slides.
For 199 skin biopsies collected at Stanford Hospital, Winetraub carried out an OCT scan before pathologists created H&E slices. He and his colleagues developed a way of putting molecular tags on the surface of the biopsies so they could be sure exactly where in the OCT scan each H&E slice came from. Then, Winetraub paired up 1,005 of these H&E images with the corresponding OCT images and entered them into an artificial intelligence algorithm that could learn how to create accurate H&Es from the raw OCT data.
“The uniqueness of this work lies in the method we developed to align OCT and H&E image pairs, letting machine-learning algorithms train on real tissue sections and providing clinicians with more accurate virtual biopsies,” Winetraub says.
The researchers fine-tuned the AI program by showing it an additional 553 pairs of H&E and OCT images before testing it out on new OCT images. When three Stanford dermatologists analyzed random assortments of true H&E images and those created from the OCT scans, they could detect cellular structures at a similar rate. Any number of H&E images can be created from a single OCT image, virtually slicing the three-dimensional reconstruction in any direction.
The study is published in Science Advances.
