Artificial Intelligence in Functional Urology: How It may Shape the Future

Imad Bentellis; Sonia Guérin; Zine-Eddine Khene; Rose Khavari; Benoit Peyronnet


Curr Opin Urol. 2021;31(4):385-390. 

In This Article

Artificial Intelligence as a Diagnostic Tool in Functional Urology

Clinical Evaluation

In a cohort of 8,071 elementary school students, Tokar et al.[12] used an ML approach to build model predicting enuresis in children. Out of 34 variables assessed, fourteen were independently associated with the occurrence of enuresis. Using a logistic regression algorithm, the authors found an accuracy of 81.3% for the prediction of enuresis. Although the authors claim that AI may prevent clinical errors due to 'human cognitive biases' one should note that the use of AI in such a setting remains restrained by the limited number of data included in the dataset used. Indeed, AI would not 'magically' point out data that are not present in a dataset and then such study is still prone to the so-called cognitive bias at the time where one selects the variables to be included in the dataset. The diagnostic relevance of such a use of AI could also be called into question, as it cannot outperform the diagnostic performances of the physicians involved in the training cohort.


Deep learning, a subset of machine learning, can be used for medical image recognition. Interpretation of dermatological images has been one of the pioneering uses of 'computer vision' in medicine and recent data suggest that ANN may outperform human dermatologists in the diagnosis of some dermatological lesions.[13] Such AI technologies can find several applications in functional urology. The role of dynamic MRI in the diagnosis and quantification of pelvic organ prolapse (POP) remains controversial especially in terms of added value compared to physical examination.[14] One of the other possible limitations of dynamic MRI may be the interobserver reproducibility of its interpretation. Two recent studies explored the interest of deep-learning-based image recognition to diagnose and grade POP on dynamic MRI.[15,16] Analyzing 15 dynamic MRI of female patients with POP, Onal et al. found that a semiautomated pelvic floor measurement algorithmic model was accurate for POP detection and quantification. Nekooeimehr and colleagues presented a method of automated contour tracking and trajectory classification of pelvic organs on dynamic MRI. They demonstrated the high reliability of their algorithmic automated model in a series of 94 cases.

As mentioned in the basic research section, AI can be used to extract texture feature on CT or MRI. In their aforementioned study, Khene et al. found statistically significant associations between some radiomic parameters and urodynamic findings. The authors concluded that CT texture analysis of the bladder wall might be an interesting tool to identify spina bifida patients with high-risk urodynamic features and could hypothetically become an alternative or a complement to urodynamic in the future in neurogenic populations.


Wang et al.[17] attempted to identify patterns of DO that correlates with clinical findings in order to minimize human interference and standardize interpretation of urodynamic tracts. They used Manifold learning[18,19] and dynamic time warping algorithms. They reported an overall accuracy of 81.35%, a sensitivity of 76.92%, and a specificity of 81.41% of their AI-based model to detect DO. Cullingsworth et al.[20] quantified frequencies and amplitudes in low amplitude rhythmic detrusor contractions (LARC) and identified a subgroup of DO patients with LARC. The model used was an automated Fast Fourier Transform algorithm[21] which achieved a 100% specificity. This accuracy was reached, whereas LARC were detected independently of the abdominal pressure traces paving the way for a possible automation of urodynamic traces interpretation. In the third promising study on the ML-based automation of urodynamics, Niederhauser et al.[22] used 'Wavelet' time-frequency analysis with AI-based algorithms to assess the possible existence of various subgroup of overactive bladder (OAB) patients according to the amplitude and frequency of nonvoiding contractions and to investigate the possibility of automatically detecting DO. They did identify several OAB subgroups and their ML-based algorithms demonstrated excellent diagnostic performances for DO detection.