Jump to: navigation, search

Computer-aided detection of prostate cancer

Prostate Cancer

Prostate cancer is the second most common cancer in men, with 10000 new cases and 2500 deaths every year in the Netherlands (Cancer Society). These numbers increase every year due to the ageing of the general population.

Screening can help reduce mortality. The results of a large multi-center ERSPC trial (Schroder et al NEJM 2009) showed that a screening program using a prostate specific antigen (PSA) blood test and transrectal ultrasound (TRUS) biopsy could reduce mortality up to 30%. However, this mortality reduction would lead to a massive amount of over-treatment in men. To save one life, 1410 healthy men would have to do a PSA test, 225 biopsied, and 48 men treated for an indolent cancer. Screening was therefore not recommended.

Prostate MRI

Prostate MRI has become very accurate in the detection of clinically significant prostate cancer. A recent study showed that compared to conventional TRUS biopsy, MR biopsy finds 18% more clinically significant cancers while reducing the number of required biopsies by a third (Pokorny et al Eur Urol 2014). A prostate MR evaluation study in biopsy naive men with a template guided biopsy reference standard showed a sensitivity of 97% for clinically significant cancers (Grey et al BJUI 2015). The negative predictive value (NPV) was 98%, which explains why a negative MR can avoid an unnecessary biopsy. It is important to realize that the majority of prostate cancers are indolent and will never become life threatening. It is however important to correctly distinguish between the indolent and clinically significant cancers. Prostate MRI is very good at this. It shows significant cancers, but may not show all indolent cancers. In contrast if a TRUS biopsy finds an indolent cancer it may still have missed a clinically significant part.

Top-Left: Diffusion weighted image. Top-Right: Apparent diffusion coefficient-map. Bottom-Left: Dynamic contrast enhanced image. Bottom-Right: T2-weighted image

Current prostate MRI consists of three different types of MR images and is referred to as multi-parametric MRI or mpMRI. Each parameter contributes to anatomical and functional information about the prostate tissue. The three typical parameters are T2-weighted, diffusion-weighted and dynamic contrast enhanced imaging. High-resolution T2-weighted images are used to assess the anatomical structure and tissue texture of the prostate. Diffusion-weighted imaging provides a biomarker for cellular tissue structure. Dynamic contrast enhanced images provide information on microvasculature and extra-cellular space.

Prostate MRI challenges

Prostate MRI requires substantial experience and good quality equipment and operators to correctly detect and categorize prostate cancer. Reported diagnostic accuracy varies strongly. Specificity was only 60% in Grey et al BJUI 2015. Another similar study Abd-Alazeez et al 2014 reported a specificity of 21%. Yet Pokorny et al Eur Urol 2014 report a specificity of 95%. To reduce variability radiologists have been setting up a consensus set of guidelines on how to acquire and report mpMPI. The second version of that reporting system has been defined in PIRADS 2.0 2015. Reducing this variability is a active research challenge.

Increasing specificity is another challenge. To substantially reduce biopsies, or even start screening, specificity is rather low, it is currently in the range of 30-50%. Yet, it is in the order of 90-99.9% in breast cancer screening.

To effectively reduce biopsies and overtreatment, MRI should be used at a large scale. The challenge is to do this cost-effectively. Challenges are many. Research in reduction of scan time is ongoing by either improving MR technology, or decreasing the minimal number of parameters. Other research focusses on the optimization of workflow by means of CAD. CAD can help optimize the information extraction, help reduce the viewing procedures, help remain vigilant in high volume reading conditions.

Research in computer-aided detection of prostate cancer

Prostate MR is a complex multi-parametric image analysis problem. This is the domain where Computer Aided Diagnosis (CAD) can help improve prostate cancer MR diagnosis. CAD's trainable pattern recognition and multi-variate statistical analysis can optimally use available information, at a constant high vigilance level.

We have been researching CAD for mpMRI for many years. We have established that CADx ( Vos 2008, Vos 2010) is capable of providing a likelihood of cancer for a region of interest in mpMRI. A CADx prototype was validated in Hambrock 2012 and was shown to help increase reader accuracy of mpMRI. Subsequent exploration lead to prototype technology capable of fully automatic detection of prostate cancer in Litjens 2014. That system was capable of finding 80% of all clinically relevant cancers at 1 false positive per mpMRI which is still state-of-the-art in the literature on CAD for mpMRI. The prototype was validated by simulation to improve reader accuracy Litjens 2015.

From Litjens 2014. Left: Segmentation of the prostate zones. Middle: Cancer probability map with ground truth in white. Right: CAD result in red with probability value, ground truth in white.

CAD for prostate cancer still has many challenges ahead and further research is ongoing in our lab and in other research groups:

  • As of yet CAD approaches the performance of a good reader, but not an expert. CAD needs to be at least as good as an expert to help readers to achieve at expert level.
  • Specificity remains an important problem. The number of biopsies is still too high and recall for screening would still be a costly issue. Learning to discriminate between lesions that are clinically relevant and the indolent or benign confounders is critical in the successful application of CAD.
  • Automatic segmentation of the prostate, surrounding anatomy and/or it's zonal structure is critical for the implementation of CAD.
  • Quantitative MRI has been shown to improve the informative value of mpMRI for both CAD and human readers. Vendor, machine setting, and patient specific variations remain a source of uncertainty that needs further research.
  • Texture analysis is increasingly being used by skilled human observers to increase their accuracy. The novel technology of deep learning can play a role in trying to discover image patterns that help discriminate relevant tissue textures.



Key publications