A new and innovative automated computer technique has been developed by CNBP researchers that is able to significantly aid in the diagnosis of bladder cancer.
The technique—which allows suspect lesion images to be quickly and effectively analysed and then classified for cancer risk, has been reported in the medical journal ‘Urologic Oncology’.
“What we’ve done is develop a computer program to carry out an automated analysis of cystoscopy images,” says lead author of the research, Dr Martin Gosnell, Researcher at the ARC Centre of Excellence for Nanoscale BioPhotonics (CNBP) at Macquarie University and Director at Quantitative Pty Ltd.
Cystoscopy is one of the most reliable methods for diagnosing bladder cancer explains Dr Gosnell.
“Images are taken of the bladder and its insides for suspicious lesions during a routine clinical patient evaluation. Dependent on the findings, this initial scan can then be followed up by a referral to a more experienced urologist, and a biopsy of the suspicious tissue can be undertaken.”
The issue says Dr Gosnell is that the clinician examining the initial images makes a visual judgement based on their professional expertise as to the next steps of action that should be undertaken—such as the need to take a biopsy for subsequent pathological analysis.
“Potential errors and unnecessary further interventions may result from the subjective character of this initial visual assessment.”
“What we’ve done,” says Dr Gosnell, “is to create an automated image analysis technique which can identify tissue and lesions as either high-risk or minimal-risk.”
Journal: Urologic Oncology.
Publication title: Computer-assisted cystoscopy diagnosis of bladder cancer.
Authors: Martin E. Gosnell (pictured top), Dmitry M. Polikarpov, Ewa M. Goldys, Andrei V. Zvyagin and David A. Gillatt.
One of the most reliable methods for diagnosing bladder cancer is cystoscopy. Depending on the findings, this may be followed by a referral to a more experienced urologist or a biopsy and histological analysis of suspicious lesion. In this work, we explore whether computer-assisted triage of cystoscopy findings can identify low-risk lesions and reduce the number of referrals or biopsies, associated complications, and costs, although reducing subjectivity of the procedure and indicating when the risk of a lesion being malignant is minimal.
Materials and methods
Cystoscopy images taken during routine clinical patient evaluation and supported by biopsy were interpreted by an expert clinician. They were further subjected to an automated image analysis developed to best capture cancer characteristics. The images were transformed and divided into segments, using a specialised color segmentation system. After the selection of a set of highly informative features, the segments were separated into 4 classes: healthy, veins, inflammation, and cancerous. The images were then classified as healthy and diseased, using a linear discriminant, the naïve Bayes, and the quadratic linear classifiers. Performance of the classifiers was measured by using receiver operation characteristic curves.
The classification system developed here, with the quadratic classifier, yielded 50% false-positive rate and zero false-negative rate, which means, that no malignant lesions would be missed by this classifier.
Based on criteria used for assessment of cystoscopy images by medical specialists and features that human visual system is less sensitive to, we developed a computer program that carries out automated analysis of cystoscopy images. Our program could be used as a triage to identify patients who do not require referral or further testing.
Below: Dr Martin Gosnell and Prof Ewa Goldys.