Urology is increasingly trending toward the incorporation of artificial intelligence (AI) into its practice, especially in imaging and pathology interpretation. 
Applying AI: computer vision, machine learning (ML) or deep learning (DL) techniques to automate the interpretation of tracings, could shorten and simplify urodynamic study review for diagnosis. 

We aim to validate the reliability of Computer Vision and Machine Learning Techniques for diagnostic aid in basic cystometric parameters by comparing the results  with those made by two expert reviewers.

We reviewed Cystometry (CMG) plots corresponding to 517 adult patients performed over a year (2023) in a single center with the same equipment and technique, according to ICS standards. Traces were anonymized for review. Pediatric patients and studies with simultaneous EMG were excluded. The diagnosis previously made by 2 expert observers showed 287 images of studies with stable detrusor and 233 images with overactive detrusor. The study focused on CMG to validate the methodology.
Images were analyzed using 2 AI techniques, with the AI system blind to the previous diagnosis.
1) Deep learning techniques based on VGG16- convolutional neural networks (CNN) architecture. To visualize decisive areas for classification, we used gradcam plus-plus.
2) Computer vision and machine learning techniques, transforming the different pressure signals from the time domain to the frequency domain using Daubechies Wavelet Transforms. This process removes noise and allows the implementation of precise thresholds for effective signal reconstruction. Sections prior to infusion onset and after maximum cystometric capacity (end of filling phase) were eliminated. To eliminate possible noise generated by empty bladder sensors, data during the first 75 ml of infusion were discarded. A critical threshold of 15 cmH2O of Pdet was applied to define involuntary contraction. This threshold enables discernment of involuntary contractions, enriching expert interpretation with quantitative data and offering a more detailed distinction between diagnostic categories. 

We defined accuracy as agreement between the diagnosis previously made by the urologist (supervised data) and the diagnosis made by the automatic system (blind to the urologist diagnosis).

Patients signed an informed consent allowing their data to be  reviewed for clinical research,  and the protocol  was approved by the regional ethics research committee  (Study  2024 06 URO CMT)

Deep learning techniques (CNN-VGG16) provided an accuracy of 75% in detecting involuntary contractions. It did not provide quantitative data (i.e. time, volume), but was able to classify  overactive detrusor and was able to focus on leakage during the CMG ( Fig.1) . 

The use of Dauebechis Wavelets yielded a diagnostic precision of 84.2%, with a specificity of 82.6% (± 4.4%) and sensitivity of 86.3% (± 4.4%). The methodology allows precise identification of the time, volume, and duration of contractions, as well as cough, leakage and artifacts.  It also enabled a more accurate bladder compliance calculation by disregarding contractions ( Figs. 2 and 3) 

During plot analysis, artifacts compromising diagnostic accuracy were detected and defined: tube movements or knocks, catheter expulsion, line flushing, and lines open to the syringe. Correction and filtering techniques were used, such as Wavelet Transforms to correct tube movements or knocks and pressure spikes due to flushing. Patient position changes artifacts do not affect Pdet reading accuracy. For cough events and artifacts due to lines open to the syringe, we used the Isolation Forest anomaly detection method. This allowed detection of changes in abdominal and vesical pressures when they coincide above a threshold within a specific time window, in which case they were marked as cough events; if an anomaly appears only in Pabd,  it is interpreted as an artifact possibly due to an open line. In 15.8% of cases there was a discrepancy between the AI system and the experts and it was mainly due to the presence of multiple artifacts or borderline values.

Each AI technique has its advantages for CMG review:  Deep learning CNN showed a satisfactory (75%) accuracy and detection of significant changes in tracings, but is not able to provide quantitative analysis.  Daubechies Wavelet has a higher accuracy (84%) in classifying graphs and analyze all the quantitative data, thus increasing interpretability. This method could save time in reviewing studies.
Our methodology emphasizes the importance of subtle differences, providing an advantage over deep learning classification approaches or classical methods such as SVM (Support Vector Machine). Authors who have used this method [2] to detect overactive detrusor (using mean, variance, median, minimum, and maximum value of each pressure signal, with a total of 15 parameters) achieved a concordance in the time domain of 62.4% ± 5.2%, and in the frequency domain using FFT, 74.0% ± 6.3%, although signal artifact corrections are not mentioned. Hobb et al. achieved better results using windowing (sensitivity 68.3% ± 5.3%, specificity 92.9% ± 1.1%) with the difficulty of having to supervise each window indicating whether an involuntary contraction occurs or not.

The integration of Wavelet Transforms and machine learning contributes to the classification of urodynamic events in CMG, allowing more accurate detection of detrusor involuntary contractions and low bladder compliance. This study surpasses traditional methods, addressing challenges imposed by common artifacts in urodynamic measurements. The application of advanced computer vision techniques and specific algorithms has proven to be fundamental in improving the objectivity and precision of evaluations. These advancements allow a detailed analysis of quantitative information in clinical practice, facilitating a semi-automatic review of graphs and enabling more reliable diagnoses and tailored treatments for lower urinary tract disorders. The combination of AI techniques with expert supervision could provide a practical system to generate high quality Urodynamic reports.

Gammie A, D'Ancona C, Kuo HC, Rosier PF. ICS teaching module: Artefacts in urodynamic pressure traces (basic module). Neurourol Urodyn. 2015 Sep 15. doi: 10.1002/nau.22881. [Epub ahead of print] Review. PubMed PMID: 26372678.
Hobbs KT, Choe N, Aksenov LI, Reyes L, Aquino W, Routh JC, Hokanson JA. Machine Learning for Urodynamic Detection of Detrusor Overactivity. Urology. 2022 Jan;159:247-254. doi: 10.1016/j.urology.2021.09.027. Epub 2021 Oct 29. PMID: 34757048; PMCID: PMC8865755.
D’Ancona C, Haylen B, Oelke M, et al. The International Continence Society (ICS) report on the terminology for adult male lower urinary tract and pelvic floor symptoms and dysfunction. Neurourology and Urodynamics. 22001199;;318–:4453.3h–4tt7p7s.: /h/dttopis.:o//rdgo/1i.0o.r1g0/1002.1/n0a0u2./2n3a8u9.273897

Continence 12S (2024) 101449
DOI: 10.1016/j.cont.2024.101449