AI in Urology: Three Recent Developments
Welcome to Health Connect, the podcast for health professionals through which we will share the latest news and information on science and technology in the medical field. In this episode, we present you with three investigations addressing the development of a deep-learning-based biomarker for the prediction of prostate cancer biochemical recurrence, a predictive model of early recovery from post-prostatectomy incontinence, and a computerized decision support system for bladder cancer treatment response assessment.
Did you know that the rise of prostate-specific antigen levels in the blood, known as biochemical recurrence, occurs in approximately 30 percent of patients after undergoing prostatectomy for prostate cancer?1
Despite its uncertain role in prostate cancer screening, prostate-specific antigen, or PSA, is a relevant biomarker in the follow-up of patients after prostatectomy. This is based on the fact that four to six weeks after successful surgery, the PSA concentration is mostly undetectable, usually around less than 0.1 nanograms per milliliter. Accordingly, increased levels of this protein, or biochemical recurrence, can signal regrowth of prostate cancer cells, this is a prognostic indicator of progression to clinical metastasis and death.1
As we know, the assessment of the risk of recurrence is assessed in clinical practice by a combination of Gleason growth patterns, also referred to as ISUP grade groups, the PSA value at diagnosis and the TNM staging criteria. Pathologists assign cancerous tissue into one of five groups, the fifth being the most aggressive tumor type.1
However, despite updates since their inception in the 1960s, Gleason patterns remain relatively limited descriptors of tissue morphology, ignoring within-grade morphologic patterns and subtle histopathologic features with vast predictive potential.1
So, could artificial intelligence hold the key to unlocking the true prognostic value of morphological assessment of cancer? To answer this question, in 2022 Pinckaers and colleagues reported a deep learning system used to develop a prognostic biomarker based on tissue morphology for early recurrence in patients with prostate cancer treated by radical prostatectomy.1
The researchers used a previous nested case-control study containing 685 patients with clinically localized prostate cancer, who received radical retropubic prostatectomy. After the procedure, the median follow-up was four years. 72 percent of the participants had biochemical recurrence.1
Four tissue microarray spots were collected from each participant. Tissue spots were cores with a diameter of 0.6 millimeters; those with less than 25 percent of tissue were discarded, and all samples were manually analyzed for prostate cancer, resulting in a total of 2,343 tissue microarrays obtained.1
In the next step, the dataset was divided, resulting in a set of 503 unique patients used for the development of the system, and a test set of 91 case-control pairs, matched for age at surgery, race, pathologic stage, and Gleason sum. Then, cross-validation was used on both sets, randomly subdividing them into three folds. Importantly, the model had multiple runs on each fold, obtaining similar results.1
To develop the deep-learning-based biomarker, convolutional neural networks were trained on the tissue sections to provide a continuous score depending on the perceived speed of recurrence. The tissue microarray spots of cases with recurrence were assigned a value from zero to four, depending on the year in which the first outcome was recorded, where zero meant recurrence within one year and four meant after more than four years.1
To validate the biomarker, the researchers used a completely independent external cohort of 204 patients with localized prostate cancer treated with radical prostatectomy; they had a median follow-up of five years. Nineteen percent of these patients had biochemical recurrence after complete remission. Samples were taken from the largest tumor focus or any higher-grade focus larger than three millimeters, with a total of 620 tissue microarray spots.1
The results show that the biomarker had a strong correlation with biochemical recurrence in the two independent cohorts of patients. In the first test set, it achieved an odds ratio of 3.32, with a p-value equal to 0.001 per unit, and matched on the Gleason sum over other factors contained in the pathology reports.1
On the other hand, in the external validation set, a hazard ratio of 3.02 was achieved with a p-value equal to 0.03 per unit, adjusted for ISUP grade, pathologic stage, preoperative prostate-specific antigen concentration, and surgical margin status. These findings support the original hypothesis of the authors about morphological information existing in the tissue beyond the ISUP grade.1
Furthermore, it is relevant to consider that the samples from both cohorts were prepared in different pathology laboratories, showing the robustness of the system with respect to differences in tissue preparation, staining protocols, and scanners.1
Another interesting feature of the study is its approach to the black box problem, inherent to deep learning methods and particularly criticized when it comes to applications in medicine, where it is essential that decisions based on an algorithm are transparent, explainable, and widely validated.1
The researchers used a modern technique called Automatic Concept Explanations, which allows visualizing the patterns learned by the network and finding concepts correlated with the patient's prognosis. Upon visual inspection, a pathologist identified concepts correlated with the speed of biochemical recurrence, the most discriminative being those that followed the morphological patterns of the Gleason classification.1
In summary, the concepts with adverse behavior showed mainly the Gleason pattern four and, to a lesser extent, Gleason pattern five, with a cribriform structure in the tissue microarrays indicating the most adverse behavior. The two intermediate concepts showed mainly stroma and less aggressive growth patterns. The two concepts categorized within slower recurrence showed mostly Gleason pattern three, with well-defined, easily recognizable glands, as compared to disorganized sheets of prostate cancer cells.1
Cápsula
Given the complicated physiology of micturition involving several organs and the nervous system, it is difficult to predict the state of continence after prostatectomy. Despite the improved surgical techniques, the incidence of post-prostatectomy incontinence remains high, particularly in the early postoperative period. Up to 20 percent of patients consistently experience this major complication for up to one year after surgery.2
Several factors may influence a patient's chances of recovery, including factors such as advanced age, high body mass index, presence of comorbidities, and history of transurethral resection of the prostate, or lower urinary tract symptoms. Anatomic features such as preoperative membranous urethral length, postoperative bladder neck preservation, and the characteristics of the anatomical pelvic structures around the urethra also play a role. In addition, the type of surgery and preservation of the neurovascular bundle, both associated with recovery rates, must equally be considered. Nevertheless, preoperative clinical data and anatomic features revealed by preoperative MRI can be used to predict early recovery from post-prostatectomy incontinence after prostatectomy.2
Fin de la cápsula.
Welcome back. In the previous segment, we discussed the prediction of biochemical recurrence, the first sign of metastatic prostate cancer after radical prostatectomy.1 Now, we will discuss the application of artificial intelligence to the prognostic of post-prostatectomy incontinence, an important functional complication of this treatment, which can negatively affect a patient's quality of life.2
As mentioned in the capsule, multiple factors influence a patient's chances of improvement, but the complex relationship among them makes it difficult to forecast the recovery from post-prostatectomy incontinence. However, since machine learning approaches allow analyzing data with numerous confounding factors and intricately related variables, the researchers Seongkeun Park and Jieun Byun developed a predictive model for early recovery from incontinence. The authors used only clinical data and anatomical features obtained during the preoperative examinations to generate the predictive model and compared its diagnostic performance with conventional logistic regression. Let's look at the evidence.2
This study retrospectively analyzed the records of 166 patients with prostate cancer who underwent prostatectomy from 2016 to 2019, and who had preoperative pelvic skeletal muscle measurements using magnetic resonance imaging. According to the date of surgery, participants were divided into a development cohort with 109 individuals to train the algorithm, and a test cohort of 57 patients to verify the method. Early recovery was defined as no use or occasional use of incontinence pads three months after surgery.2
Among the baseline characteristics, the mean age of the participants was 71.6 years, and incontinence recovery was achieved in 51.4 percent of cases within three months after prostatectomy in the development group, and in 35.1 percent of the test group. The surgical methods and magnetic resonance imaging findings between the two groups were significantly different.2
Let’s examine these differences. In characterizing the patients with early recovery, the researchers found that they were significantly younger than those with consistent incontinence, with a mean of 70.1 versus 72.8 years of age. Magnetic resonance imaging findings showed that the membranous urethral length of those who recovered was greater, at 15.7 millimeters versus 13.9 millimeters, and the obturator internal muscle was thicker, at 8.2 millimeters versus 17.5 millimeters compared to patients that developed incontinence.2
After univariate logistic regression analysis, age and membranous urethral length were significantly related to recovery within three months. In multivariate logistic regression analysis, both factors remained significant independent predictors of recovery, with age achieving an odds ratio of 1.07 and a p-value equal to 0.007, while membranous urethral length achieved an odds ratio of 0.87, with a p-value equal to 0.002.2
Based on these results, an analysis of the traditional logistic regression model showed low to moderate diagnostic accuracy in predicting early recovery, with an area under the curve of 0.59. In comparison, prediction models built using machine learning algorithms, namely k-nearest neighbor, decision tree, support vector machine, and random forest, achieved an improved diagnostic performance, with a higher area under the curve compared to the logistic regression model. This was especially notable in the support-vector machine algorithm, which achieved an area under the curve of 0.65.2
To recapitulate, one of the most relevant findings of this study is that the probability of continence recovery within three months after prostatectomy increased with membranous urethral length and decreased with age. This could be explained in light of the finding by Dubbelman and investigators in 2012, who observed a 27 percent decrease in maximal urethral closure after prostatectomy. Taken together, this tells us that patients with a longer membranous urethral length are more likely to maintain sphincter function. On the other hand, it is known that with aging there is usually a general loss of the ability to control micturition. In 2009, Greco and his team found that age younger than 70 years was a highly relevant factor for continence recovery within three months after surgery.2
The findings of Park and Byun, published in 2021, provide evidence that machine learning algorithms can achieve higher diagnostic accuracy compared to conventional statistical approaches. Talking about clinical applications, the determination of preoperative factors that may characterize the patient at high risk of incontinence would be valuable for clinicians in consulting patients before surgery, and in explaining the postoperative recovery process. At the same time, this tool may provide valuable information for treatment, allowing conservative treatments to be administered at an early stage to high-risk patients, and avoiding unnecessary aggressive surgical correction for patients who can expect recovery.2
Now, let’s move on to a small multi-institutional observational study, led by researcher Di Sun and published in 2022. The study proposes a computerized artificial intelligence-based decision support system to improve physicians' accuracy in assessing bladder cancer patients' response to chemotherapy before radical cystectomy.3
To provide some context, in 2021 close to 83,730 new cases of bladder cancer would be diagnosed and 17,200 patients would die from it in the United States alone, with a five-year relative survival rate of 77 percent for all SEER stages combined, an acronym that stands for Surveillance, Epidemiology, and End Results.3
As we know, early diagnosis and treatment can improve the survival rate; in particular, neoadjuvant chemotherapy before radical cystectomy can decrease the likelihood of metastatic disease. However, it is essential to evaluate the response of bladder lesions to chemotherapeutic treatment to avoid unnecessary toxicity or to support an organ-preserving therapy.3
The authors' previous work showed that a computerized decision support system for bladder cancer treatment response assessment based on artificial intelligence, or CDSS-T for short, can improve the performance of physicians from a single institution. The present study aimed to further investigate the impact of this tool on the performance of physicians from different specialties and institutions.3
To that end, pre- and post-chemotherapy computed tomography urography scans were obtained from 123 patients, with a total of 157 pairs of cancers before and after treatment. Information on the pathological stage of cancer after treatment was collected to determine whether a patient had a complete response as a reference standard, indicating that 40 lesion pairs had a T0 stage after chemotherapy, and 117 lesion pairs had an incomplete response after chemotherapy, indicated as greater than T0.3
The cancers were then segmented on the computed tomography urography images and multiple regions of interest were extracted from these slices to train a system, which combined a deep learning convolutional neural network and a radiomics model. The aim was to estimate the probability of response to neoadjuvant chemotherapy treatment, producing a discriminant score in the range of one to 10, where a lower number indicated a lower probability that the post-treatment lesion had a complete response.3
Seventeen physicians participated as observers, including abdominal radiologists, a urologist, and oncologists with experience in assessing response to bladder cancer treatment, as well as an inexperienced diagnostic radiology resident, a neurology fellow, and a medical student.3
Each physician provided a response assessment by reading the pre-and-post-treatment scan pair for each cancer. First, without CDSS-T, an estimate of the percentage response to treatment was given on a scale of minus 100 percent to plus 100 percent using the Response Evaluation Criteria in Solid Tumors, known as RECIST criteria. Using this estimate, zero percent indicated no change between the pre-and post-chemotherapy scan, and 100 percent indicated a complete response. Second, the physician assessed the tumor response category, also based on the RECIST criteria. Third, the probability of reaching stage T0 was estimated on a scale of zero to 100, where zero indicates no chance of the cancer being at stage T0 after treatment. Finally, a recommendation was given for the next step of treatment, which included surgery and radiation.3
For the next step of the study, physicians were shown the score calculated by the model, allowing them to revise their ratings or leave them unchanged. Intraobserver variability was estimated from repeated readings in 51 randomly sampled cases. In addition, the level of difficulty in assessing the response of each cancer was evaluated using the standard deviation of the participating radiologists' estimates for the probability of stage T0.3
Thus, the assessment was classified as relatively easy when the standard deviation was less than 25, or relatively difficult when the standard deviation was greater than 25. There were 95 pairs of pre-post-treatment scans in the easy subset, with 77 cancers that did not fully correspond labeled as greater than T0; while the remaining 62 pairs were grouped in the difficult subset, with 40 cancers that did not fully respond.3
Now that we have discussed the methodology, what were the most relevant findings? First, the average performance of the 17 observers in assessing response to bladder treatment was significantly improved when aided by the model. While the diagnostic difficulty of the cancer cases had an impact on the original performance, there was an increase in the accuracy with the aid of CDSS-T: the average area under the curve without the help of the system was 0.80 for the easy subset and 0.58 for the difficult subset, while with CDSS-T, the area under the curve increased to 0.84 in the easy subset and 0.62 in the difficult subset.3
Furthermore, physician experience contributed only a small difference in the level of improvement in the assessment of response to cancer treatment. Without CDSS-T, the performance of the experienced observers was better than that of the inexperienced observers. Although the experienced physicians had a significantly smaller gain in the area under the curve, this could be due to their being more confident and persistent in their own evaluations.3
More importantly, with CDSS-T, both experienced and inexperienced observers achieved the same mean area under the curve of 0.77. Therefore, the CDSS-T was able to help inexperienced clinicians to perform diagnostically at a level comparable to that of the experienced observers.3
Moreover, with the help of the CDSS-T, the physicians had less variability and better agreement in both the original evaluations and the randomly repeated assessments. On average, oncologists had a greater gain in accuracy than radiologists, suggesting that the model could be a potentially useful tool for physicians who are not radiologists, at any level of experience, to achieve greater accuracy in assessing response to bladder cancer treatment, resulting in more consistent performance among all doctors.3
Thanks for joining us on this episode of Health Connect. Don't miss our next episode, where we will discuss more artificial intelligence developments in other medical fields. Subscribe to our channel to discover the latest medical news.
References:
- 1. Pinckaers H, van Ipenburg J, Melamed J, De Marzo A, Platz EA, van Ginneken B, et al. Predicting biochemical recurrence of prostate cancer with artificial intelligence. Commun Med (Lond). 2022;2:64. Available at: https://www.nature.com/articles/s43856-022-00126-3
- 2. Park S, Byun J. A Study of Predictive Models for Early Outcomes of Post-Prostatectomy Incontinence: Machine Learning Approach vs. Logistic Regression Analysis Approach. Appl. Sci. 2021;11(13):6225. Available at: https://www.mdpi.com/2076-3417/11/13/6225
- 3. Sun D, Hadjiiski L, Alva A, Zakharia Y, Joshi M, Chan HP, et al. Computerized Decision Support for Bladder Cancer Treatment Response Assessment in CT Urography: Effect on Diagnostic Accuracy in Multi-Institution Multi-Specialty Study. Tomography. 2022;8(2):644-656. Available at: https://www.mdpi.com/2379-139X/8/2/54
Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.
NON-2022-14759
NON-2023-2512