OCT image interpretation algorithms detect bladder cancer automatically

David L. Shenkenberg

Bladder cancer currently is diagnosed by visualizing the bladder with an endoscope and taking a biopsy for pathological evaluation. Although this provides a definitive diagnosis, some malignant tissue often is missed during the process. Because optical coherence tomography (OCT) can resolve structures 15 μm or smaller, it could detect suspicious areas missed by endoscopy.

It could do so noninvasively as well, because, even though light scattering by skin limits the penetration depth of noninvasive OCT to 1 to 2 mm, most bladder cancers are ≤600 μm below the skin. Past clinical studies have shown that the technique has high sensitivity and specificity.

Now researchers from George Washington University in Washington and from Cytogen Corp. in Princeton, N.J., have developed algorithms, or sets of computer instructions, designed to automate and facilitate the interpretation of OCT scans of the bladder. They envision that the technique will be adopted by the medical community most readily as a method of guiding the endoscopist in real time.

OCT can distinguish between a healthy bladder (left) and invasive cancer (right).

The investigators used 182 OCT scans, each 200 × 200 pixels, and corresponding pathology results from a previous study at their institution. The scans were taken in 1.5 s with a 980-nm, 10-mW superluminescent LED coupled to an optical fiber in the endoscope sheath.

They instructed the computer to analyze the texture of the OCT scans by the probability at which pixel characteristics occur together. Then they used a decision tree on a training data set to correlate these features of the tissue with three groups: invasive cancer and papillary lesions, noninvasive cancer and precancerous lesions, and benign tissue. After training, they tested the algorithms on an independent data set.

As reported in the March/April 2008 issue of the Journal of Biomedical Optics, the researchers’ computer instructions had 92 percent sensitivity and 62 percent specificity and detected precancerous tissue and noninvasive cancer most accurately. The scientists plan to improve the specificity, especially for detecting invasive cancer, to develop the algorithms to identify inflammation and to test their algorithms on additional independent data sets.