Deep-learning algorithm developed for bladder Ca detection

May 22, 2019

New data suggest it may be possible to use computer-augmented cystoscopy to aid in diagnostic decision-making and improve the diagnostic yield of papillary bladder cancers.

New data suggest it may be possible to use computer-augmented cystoscopy to aid in diagnostic decision-making and improve the diagnostic yield of papillary bladder cancers.

At the AUA annual meeting in Chicago, researchers reported that they have created a deep-learning algorithm that may more accurately detect bladder cancer.

“We have demonstrated that computer-augmented cystoscopy based on a deep-learning algorithm can detect bladder tumors with high sensitivity and specificity and may serve as a new adjunct imaging technology for bladder cancer detection,” said study investigator Joseph C. Liao, MD, associate professor of urology at Stanford University School of Medicine and chief of urology at the VA Palo Alto Health Care System in Palo Alto, CA.

Also see: Germline variants found in 18% of patients with familial UC

Dr. Liao and his colleagues have developed a deep-learning algorithm using recorded videos derived from office-based cystoscopy and transurethral resection of bladder tumor (TURBT) from 100 subjects and 141 videos. Video frames containing histologically confirmed papillary bladder cancer were first manually annotated.

Using an image analysis platform called TUMNet, the authors then evaluated the videos in two stages. First, they used it to recognize frames containing abnormal areas and then segmented them within the tumor. By examining 417 cancers and 2,335 normal frames, a training set was constructed based on 95 subjects and validated in five subjects.

Next: Technology accurately detects tumors in test cohortTechnology accurately detects tumors in test cohort

The authors found that the TUMNet per-frame sensitivity was 88% and per-tumor sensitivity was 90% with a per-frame specificity of 99%. In addition, TUMNet was able to accurately detect all 16 tumors that were resected in the ongoing prospective test cohort (15 cancerous and one benign).

“We believe our deep-learning algorithm holds significant promise for clinical translation in both clinic and OR settings. Potential applications include post-hoc quality control review and real-time integration during cystoscopy and TURBT,” first author Eugene Shkolyar, MD, urology resident at Stanford University School of Medicine, told Urology Times.

In order to enable real-time integration, he and his colleagues are actively working on further streamlining the algorithm and adding automated reporting features. They are also investigating this approach for identifying flat tumors, including carcinoma in situ and benign lesions.

Read: Routine PET/CT not beneficial for node-negative patients

“Prior to disseminating the technology, we aim to fully evaluate its clinical utility, and are hoping to conduct studies evaluating its use both real-time and for recorded videos,” said Dr. Shkolyar.

He noted that more than 1 million cystoscopies are performed annually in the U.S. for detection and surveillance of bladder cancer. Yet, studies suggest that standard cystoscopy may fail to detect up to 20% of bladder cancers. Using deep-learning algorithms as diagnostic tools during endoscopy in other fields has been shown to improve care by providing additional quality control and standardization.

Dr. Liao said with bladder cancer, the diagnostic accuracy of cystoscopy depends on provider experience and ability to recognize a variety of benign and cancerous lesions in the bladder.

“Computer-aided cystoscopy can help to reduce the variability between providers by serving as a second observer. It can also provide guidance for trainees and physician extenders in underserved areas where access to urologists is limited,” said Dr. Liao.

Yair Lotan, MD, professor of urology and chief of urologic oncology at UT Southwestern Medical Center, Dallas, said this approach sounds like it has potential, but much more validation will be required. He said the deep-learning algorithm will need to be much more thoroughly investigated to understand its strengths and weaknesses.

“It sounds really cool. We have to wait to see if it is clinically useful,” said Dr. Lotan, who was not involved with the research.