Addition of inflammation-related biomarkers to the CAIDE model for risk prediction of all-cause dementia, Alzheimer's disease and vascular dementia in a prospective study.

BACKGROUND: It is of interest whether inflammatory biomarkers can improve dementia prediction models, such as the widely used Cardiovascular Risk Factors, Aging and Dementia (CAIDE) model. METHODS: The Olink Target 96 Inflammation panel was assessed in a nested case-cohort design within a large, population-based German cohort study (n = 9940; age-range: 50-75 years). All study participants who developed dementia over 20 years of follow-up and had complete CAIDE variable data (n = 562, including 173 Alzheimer's disease (AD) and 199 vascular dementia (VD) cases) as well as n = 1,356 controls were selected for measurements. 69 inflammation-related biomarkers were eligible for use. LASSO logistic regression and bootstrapping were utilized to select relevant biomarkers and determine areas under the curve (AUCs). RESULTS: The CAIDE model 2 (including Apolipoprotein E (APOE) ε4 carrier status) predicted all-cause dementia, AD, and VD better than CAIDE model 1 (without APOE ε4) with AUCs of 0.725, 0.752 and 0.707, respectively. Although 20, 7, and 4 inflammation-related biomarkers were selected by LASSO regression to improve CAIDE model 2, the AUCs did not increase markedly. CAIDE models 1 and 2 generally performed better in mid-life (50-64 years) than in late-life (65-75 years) sub-samples of our cohort, but again, inflammation-related biomarkers did not improve their predictive abilities. CONCLUSIONS: Despite a lack of improvement in dementia risk prediction, the selected inflammation-related biomarkers were significantly associated with dementia outcomes and may serve as a starting point to further elucidate the pathogenesis of dementia.

Metrics reloaded: recommendations for image analysis validation.

Increasing evidence shows that flaws in machine learning (ML) algorithm validation are an underestimated global problem. In biomedical image analysis, chosen performance metrics often do not reflect the domain interest, and thus fail to adequately measure scientific progress and hinder translation of ML techniques into practice. To overcome this, we created Metrics Reloaded, a comprehensive framework guiding researchers in the problem-aware selection of metrics. Developed by a large international consortium in a multistage Delphi process, it is based on the novel concept of a problem fingerprint-a structured representation of the given problem that captures all aspects that are relevant for metric selection, from the domain interest to the properties of the target structure(s), dataset and algorithm output. On the basis of the problem fingerprint, users are guided through the process of choosing and applying appropriate validation metrics while being made aware of potential pitfalls. Metrics Reloaded targets image analysis problems that can be interpreted as classification tasks at image, object or pixel level, namely image-level classification, object detection, semantic segmentation and instance segmentation tasks. To improve the user experience, we implemented the framework in the Metrics Reloaded online tool. Following the convergence of ML methodology across application domains, Metrics Reloaded fosters the convergence of validation methodology. Its applicability is demonstrated for various biomedical use cases.

Robust incorporation of historical information with known type I error rate inflation.

Bayesian clinical trials can benefit from available historical information through the specification of informative prior distributions. Concerns are however often raised about the potential for prior-data conflict and the impact of Bayes test decisions on frequentist operating characteristics, with particular attention being assigned to inflation of type I error (TIE) rates. This motivates the development of principled borrowing mechanisms, that strike a balance between frequentist and Bayesian decisions. Ideally, the trust assigned to historical information defines the degree of robustness to prior-data conflict one is willing to sacrifice. However, such relationship is often not directly available when explicitly considering inflation of TIE rates. We build on available literature relating frequentist and Bayesian test decisions, and investigate a rationale for inflation of TIE rate which explicitly and linearly relates the amount of borrowing and the amount of TIE rate inflation in one-arm studies. A novel dynamic borrowing mechanism tailored to hypothesis testing is additionally proposed. We show that, while dynamic borrowing prevents the possibility to obtain a simple closed-form TIE rate computation, an explicit upper bound can still be enforced. Connections with the robust mixture prior approach, particularly in relation to the choice of the mixture weight and robust component, are made. Simulations are performed to show the properties of the approach for normal and binomial outcomes, and an exemplary application is demonstrated in a case study.

Simulating and reporting frequentist operating characteristics of clinical trials that borrow external information: Towards a fair comparison in case of one-arm and hybrid control two-arm trials.

Borrowing information from historical or external data to inform inference in a current trial is an expanding field in the era of precision medicine, where trials are often performed in small patient cohorts for practical or ethical reasons. Even though methods proposed for borrowing from external data are mainly based on Bayesian approaches that incorporate external information into the prior for the current analysis, frequentist operating characteristics of the analysis strategy are often of interest. In particular, type I error rate and power at a prespecified point alternative are the focus. We propose a procedure to investigate and report the frequentist operating characteristics in this context. The approach evaluates type I error rate of the test with borrowing from external data and calibrates the test without borrowing to this type I error rate. On this basis, a fair comparison of power between the test with and without borrowing is achieved. We show that no power gains are possible in one-sided one-arm and two-arm hybrid control trials with normal endpoint, a finding proven in general before. We prove that in one-arm fixed-borrowing situations, unconditional power (i.e., when external data is random) is reduced. The Empirical Bayes power prior approach that dynamically borrows information according to the similarity of current and external data avoids the exorbitant type I error inflation occurring with fixed borrowing. In the hybrid control two-arm trial we observe power reductions as compared to the test calibrated to borrowing that increase when considering unconditional power.

Beyond rankings: Learning (more) from algorithm validation.

Challenges have become the state-of-the-art approach to benchmark image analysis algorithms in a comparative manner. While the validation on identical data sets was a great step forward, results analysis is often restricted to pure ranking tables, leaving relevant questions unanswered. Specifically, little effort has been put into the systematic investigation on what characterizes images in which state-of-the-art algorithms fail. To address this gap in the literature, we (1) present a statistical framework for learning from challenges and (2) instantiate it for the specific task of instrument instance segmentation in laparoscopic videos. Our framework relies on the semantic meta data annotation of images, which serves as foundation for a General Linear Mixed Models (GLMM) analysis. Based on 51,542 meta data annotations performed on 2,728 images, we applied our approach to the results of the Robust Medical Instrument Segmentation Challenge (ROBUST-MIS) challenge 2019 and revealed underexposure, motion and occlusion of instruments as well as the presence of smoke or other objects in the background as major sources of algorithm failure. Our subsequent method development, tailored to the specific remaining issues, yielded a deep learning model with state-of-the-art overall performance and specific strengths in the processing of images in which previous methods tended to fail. Due to the objectivity and generic applicability of our approach, it could become a valuable tool for validation in the field of medical image analysis and beyond.

Spectral imaging enables contrast agent-free real-time ischemia monitoring in laparoscopic surgery.

Laparoscopic surgery has evolved as a key technique for cancer diagnosis and therapy. While characterization of the tissue perfusion is crucial in various procedures, such as partial nephrectomy, doing so by means of visual inspection remains highly challenging. We developed a laparoscopic real-time multispectral imaging system featuring a compact and lightweight multispectral camera and the possibility to complement the conventional surgical view of the patient with functional information at a video rate of 25 Hz. To enable contrast agent-free ischemia monitoring during laparoscopic partial nephrectomy, we phrase the problem of ischemia detection as an out-of-distribution detection problem that does not rely on data from any other patient and uses an ensemble of invertible neural networks at its core. An in-human trial demonstrates the feasibility of our approach and highlights the potential of spectral imaging combined with advanced deep learning-based analysis tools for fast, efficient, reliable, and safe functional laparoscopic imaging.

Bovine meat and milk factor protein expression in tumor-free mucosa of colorectal cancer patients coincides with macrophages and might interfere with patient survival.

Bovine milk and meat factors (BMMFs) are plasmid-like DNA molecules isolated from bovine milk and serum, as well as the peritumor of colorectal cancer (CRC) patients. BMMFs have been proposed as zoonotic infectious agents and drivers of indirect carcinogenesis of CRC, inducing chronic tissue inflammation, radical formation and increased levels of DNA damage. Data on expression of BMMFs in large clinical cohorts to test an association with co-markers and clinical parameters were not previously available and were therefore assessed in this study. Tissue sections with paired tumor-adjacent mucosa and tumor tissues of CRC patients [individual cohorts and tissue microarrays (TMAs) (n = 246)], low-/high-grade dysplasia (LGD/HGD) and mucosa of healthy donors were used for immunohistochemical quantification of the expression of BMMF replication protein (Rep) and CD68/CD163 (macrophages) by co-immunofluorescence microscopy and immunohistochemical scoring (TMA). Rep was expressed in the tumor-adjacent mucosa of 99% of CRC patients (TMA), was histologically associated with CD68+ /CD163+ macrophages and was increased in CRC patients when compared to healthy controls. Tumor tissues showed only low stromal Rep expression. Rep was expressed in LGD and less in HGD but was strongly expressed in LGD/HGD-adjacent tissues. Albeit not reaching statistical significance, incidence curves for CRC-specific death were increased for higher Rep expression (TMA), with high tumor-adjacent Rep expression being linked to the highest incidence of death. BMMF Rep expression might represent a marker and early risk factor for CRC. The correlation between Rep and CD68 expression supports a previous hypothesis that BMMF-specific inflammatory regulations, including macrophages, are involved in the pathogenesis of CRC.

Detection of Clinically Significant Prostate Cancer Using Targeted Biopsy with Four Cores Versus Target Saturation Biopsy with Nine Cores in Transperineal Prostate Fusion Biopsy: A Prospective Randomized Trial.

BACKGROUND: Multiparametric magnetic resonance imaging (mpMRI) and targeted biopsy (TB) facilitate accurate detection of clinically significant prostate cancer (csPC). However, it remains unclear how targeted cores should be applied for accurate diagnosis of csPC. OBJECTIVE: To assess csPC detection rates for two target-directed MRI/transrectal ultrasonography (TRUS) fusion biopsy approaches, conventional TB and target saturation biopsy (TS). DESIGN, SETTING, AND PARTICIPANTS: This was a prospective single-center study of outcomes for transperineal MRI/TRUS fusion biopsies for 170 men. Half of the men (n = 85) were randomized to conventional TB with four cores per lesion and half (n = 85) to TS with nine cores. Biopsies were performed by three experienced board-certified urologists. OUTCOME MEASUREMENTS AND STATISTICAL ANALYSIS: PC and csPC (International Society of Urological Pathology grade group ≥2) detection rates for systematic biopsy (SB), TB, and TS were analyzed using McNemar's test for intrapatient comparisons and Fisher's exact test for TS versus TB. A combination of targeted biopsy (TS or TB) and SB served as the reference. RESULTS AND LIMITATIONS: According to the reference, csPC was diagnosed for 57 men in the TS group and 36 men in the TB group. Of these, TS detected 57/57 csPC cases and TB detected 33/36 csPC cases (p = 0.058). Detection of Gleason grade group 1 disease was 10/12 cases with TS and 8/17 cases with TB (p = 0.055). In addition, TS detected 97% of 63 csPC lesions, compared to 86% with TB (p = 0.1). Limitations include the single-center design, the limited generalizability owing to the transperineal biopsy route, the lack of central review of pathology and radical prostatectomy correlation, and uneven distributions of csPC prevalence, Prostate Imaging-Reporting and Data System (PI-RADS) 5 lesions, men with two or more PI-RADS ≥3 lesions, and prostate-specific antigen density between the groups, which may have affected the results. CONCLUSIONS: In our study, rates of csPC detection did not significantly differ between TS and TB. PATIENT SUMMARY: In this study, we investigated two targeted approaches for taking prostate biopsy samples after observation of suspicious lesions on prostate scans. We found that the rates of detection of prostate cancer did not significantly differ between the two approaches.

The Medical Segmentation Decathlon.

International challenges have become the de facto standard for comparative assessment of image analysis algorithms. Although segmentation is the most widely investigated medical image processing task, the various challenges have been organized to focus only on specific clinical tasks. We organized the Medical Segmentation Decathlon (MSD)-a biomedical image analysis challenge, in which algorithms compete in a multitude of both tasks and modalities to investigate the hypothesis that a method capable of performing well on multiple tasks will generalize well to a previously unseen task and potentially outperform a custom-designed solution. MSD results confirmed this hypothesis, moreover, MSD winner continued generalizing well to a wide range of other clinical problems for the next two years. Three main conclusions can be drawn from this study: (1) state-of-the-art image segmentation algorithms generalize well when retrained on unseen tasks; (2) consistent algorithmic performance across multiple tasks is a strong surrogate of algorithmic generalizability; (3) the training of accurate AI segmentation models is now commoditized to scientists that are not versed in AI model training.

Spectral organ fingerprints for machine learning-based intraoperative tissue classification with hyperspectral imaging in a porcine model.

Visual discrimination of tissue during surgery can be challenging since different tissues appear similar to the human eye. Hyperspectral imaging (HSI) removes this limitation by associating each pixel with high-dimensional spectral information. While previous work has shown its general potential to discriminate tissue, clinical translation has been limited due to the method's current lack of robustness and generalizability. Specifically, the scientific community is lacking a comprehensive spectral tissue atlas, and it is unknown whether variability in spectral reflectance is primarily explained by tissue type rather than the recorded individual or specific acquisition conditions. The contribution of this work is threefold: (1) Based on an annotated medical HSI data set (9059 images from 46 pigs), we present a tissue atlas featuring spectral fingerprints of 20 different porcine organs and tissue types. (2) Using the principle of mixed model analysis, we show that the greatest source of variability related to HSI images is the organ under observation. (3) We show that HSI-based fully-automatic tissue differentiation of 20 organ classes with deep neural networks is possible with high accuracy (> 95%). We conclude from our study that automatic tissue discrimination based on HSI data is feasible and could thus aid in intraoperative decisionmaking and pave the way for context-aware computer-assisted surgery systems and autonomous robotics.

Three-dimensional Magnetic Resonance Imaging-based Printed Models of Prostate Anatomy and Targeted Biopsy-proven Index Tumor to Facilitate Patient-tailored Radical Prostatectomy-A Feasibility Study.

In this prospective single-center feasibility study, we demonstrate that the use of three-dimensional (3D)-printed prostate models support nerve-sparing radical prostatectomy (RP) and intraoperative frozen sectioning (IFS) in ten men suffering from intermediate- and high-risk prostate cancer (PC), of whom seven harbored pT3 disease. Patient-specific 3D resin models were printed based on preoperative multiparametric magnetic resonance imaging (mpMRI) to provide an exact 3D impression of significant tumor lesions. RP and IFS were planned in a patient-tailored fashion. The 36-region Prostate Imaging Reporting and Data System (PI-RADS) v2.0 scheme was used to compare the MRI/3D print with whole-mount histopathology. In all cases, localization of the index lesion was correctly displayed by MRI and the 3D model. Localization of significant PC lesions correlated significantly (Pearson`s correlation coefficient of 0.88; p < 0.001). In addition, a significant correlation of the width, length, and volume of the tumor and prostate gland, derived from the printed model and histopathology, was found, using Pearson's correlation analyses and Bland-Altman plots. In conclusion, 3D-printed prostate models correlate well with final pathology and can be used to tailor RP. PATIENT SUMMARY: The use of three-dimensional (3D)-printed prostate models based on preoperative magnetic resonance imaging (MRI) may improve prostatectomy outcome. This study confirmed the accuracy of 3D-printed prostates compared with pathology from radical prostatectomy specimens. Thus, MRI-derived 3D-printed prostate models can assist in prostate cancer surgery.

A decision-theoretic approach to Bayesian clinical trial design and evaluation of robustness to prior-data conflict.

Bayesian clinical trials allow taking advantage of relevant external information through the elicitation of prior distributions, which influence Bayesian posterior parameter estimates and test decisions. However, incorporation of historical information can have harmful consequences on the trial's frequentist (conditional) operating characteristics in case of inconsistency between prior information and the newly collected data. A compromise between meaningful incorporation of historical information and strict control of frequentist error rates is therefore often sought. Our aim is thus to review and investigate the rationale and consequences of different approaches to relaxing strict frequentist control of error rates from a Bayesian decision-theoretic viewpoint. In particular, we define an integrated risk which incorporates losses arising from testing, estimation, and sampling. A weighted combination of the integrated risk addends arising from testing and estimation allows moving smoothly between these two targets. Furthermore, we explore different possible elicitations of the test error costs, leading to test decisions based either on posterior probabilities, or solely on Bayes factors. Sensitivity analyses are performed following the convention which makes a distinction between the prior of the data-generating process, and the analysis prior adopted to fit the data. Simulation in the case of normal and binomial outcomes and an application to a one-arm proof-of-concept trial, exemplify how such analysis can be conducted to explore sensitivity of the integrated risk, the operating characteristics, and the optimal sample size, to prior-data conflict. Robust analysis prior specifications, which gradually discount potentially conflicting prior information, are also included for comparison. Guidance with respect to cost elicitation, particularly in the context of a Phase II proof-of-concept trial, is provided.

On the feasibility of pediatric dose-finding trials in small samples with information from a preceding trial in adults.

We consider the case of pediatric dose-finding trials with extremely limited sample size. The operating characteristics of the standard design, the Continual Reassessment Method (CRM), are only well described for sample sizes of about 20 patients or more. In this simulation study, we assume the situation of a pediatric trial with only 10 patients and a preceding dose-finding trial in adults. Based on the adult data, we reduce the set of pediatric doses and formulate (partially) informative prior distributions for the pediatric trial. Our simulations show that such small pediatric dose-finding trials with robustified priors may provide sufficient operating characteristics.

Improvement of PI-RADS-dependent prostate cancer classification by quantitative image assessment using radiomics or mean ADC.

Background Currently, interpretation of prostate MRI is performed qualitatively. Quantitative assessment of the mean apparent diffusion coefficient (mADC) is promising to improve diagnostic accuracy while radiomic machine learning (RML) allows to probe complex parameter spaces to identify the most promising multi-parametric models. We have previously developed quantitative RML and ADC classifiers for prediction of clinically significant prostate cancer (sPC) from prostate MRI, however these have not been combined with radiologist PI-RADS assessment. Purpose To propose and evaluate diagnostic algorithms combining quantitative ADC or RML and qualitative PI-RADS assessment for prediction of sPC. Methods and population The previously published quantitative models (RML and mADC) were utilized to construct four algorithms: 1) Down(ADC) and 2) Down(RML): clinically detected PI-RADS positive prostate lesions (defined as either PI-RADS≥3 or ≥4) were downgraded to MRI negative upon negative quantitative assessment; and 3) Up(ADC) and 4) Up(RML): MRI-negative lesions were upgraded to MRI-positive upon positive assessment of quantitative parameters. Analyses were performed at the individual lesion level and the patient level in 133 consecutive patients with suspicion for clinically significant prostate cancer (sPC, International Society of Urological Pathology (ISUP) grade group≥2), the test set subcohort of a previously published patient population. McNemar test was used to compare differences in sensitivity, specificity and accuracy. Differences between lesions of different prostate zones were assessed using ANOVA. Reduction in false positive assessments was assessed as ratios. Results Compared to clinical assessment at the PI-RADS≥4 cut-off alone, algorithms Down(ADC/RML) improved specificity from 43% to 65% (p = 0.001)/62% (p = 0.003), while sensitivity did not change significantly at 89% compared to 87% (p = 1.0)/89% (unchanged) on the patient level. Reduction of false positive lesions was 50% [26/52] in the PZ and 53% [15/28] in the TZ. Algorithms Up(ADC/RML) led, on a patient basis, to an unfavorable loss of specificity from 43% to 30% (p = 0.039)/32% (p = 0.106), with insignificant increase of sensitivity from 89% to 96%/96% (both p = 1.0). Compared to clinical assessment at the PI-RADS≥3 cut-off alone, similar results were observed for Down(ADC) with significantly increased specificity from 2% to 23% (p < 0.001) and unchanged sensitivity on the lesion level; patient level specificity increased only non-significantly. Conclusion Downgrading PI-RADS≥3 and ≥ 4 lesions based on quantitative mADC measurements or RML classifiers can increase diagnostic accuracy by enhancing specificity and preserving sensitivity for detection of sPC and reduce false positives.

Comparative validation of multi-instance instrument segmentation in endoscopy: Results of the ROBUST-MIS 2019 challenge.

Intraoperative tracking of laparoscopic instruments is often a prerequisite for computer and robotic-assisted interventions. While numerous methods for detecting, segmenting and tracking of medical instruments based on endoscopic video images have been proposed in the literature, key limitations remain to be addressed: Firstly, robustness, that is, the reliable performance of state-of-the-art methods when run on challenging images (e.g. in the presence of blood, smoke or motion artifacts). Secondly, generalization; algorithms trained for a specific intervention in a specific hospital should generalize to other interventions or institutions. In an effort to promote solutions for these limitations, we organized the Robust Medical Instrument Segmentation (ROBUST-MIS) challenge as an international benchmarking competition with a specific focus on the robustness and generalization capabilities of algorithms. For the first time in the field of endoscopic image processing, our challenge included a task on binary segmentation and also addressed multi-instance detection and segmentation. The challenge was based on a surgical data set comprising 10,040 annotated images acquired from a total of 30 surgical procedures from three different types of surgery. The validation of the competing methods for the three tasks (binary segmentation, multi-instance detection and multi-instance segmentation) was performed in three different stages with an increasing domain gap between the training and the test data. The results confirm the initial hypothesis, namely that algorithm performance degrades with an increasing domain gap. While the average detection and segmentation quality of the best-performing algorithms is high, future research should concentrate on detection and segmentation of small, crossing, moving and transparent instrument(s) (parts).

Methods and open-source toolkit for analyzing and visualizing challenge results.

Grand challenges have become the de facto standard for benchmarking image analysis algorithms. While the number of these international competitions is steadily increasing, surprisingly little effort has been invested in ensuring high quality design, execution and reporting for these international competitions. Specifically, results analysis and visualization in the event of uncertainties have been given almost no attention in the literature. Given these shortcomings, the contribution of this paper is two-fold: (1) we present a set of methods to comprehensively analyze and visualize the results of single-task and multi-task challenges and apply them to a number of simulated and real-life challenges to demonstrate their specific strengths and weaknesses; (2) we release the open-source framework challengeR as part of this work to enable fast and wide adoption of the methodology proposed in this paper. Our approach offers an intuitive way to gain important insights into the relative and absolute performance of algorithms, which cannot be revealed by commonly applied visualization techniques. This is demonstrated by the experiments performed in the specific context of biomedical image analysis challenges. Our framework could thus become an important tool for analyzing and visualizing challenge results in the field of biomedical image analysis and beyond.

Detection of Significant Prostate Cancer Using Target Saturation in Transperineal Magnetic Resonance Imaging/Transrectal Ultrasonography-fusion Biopsy.

BACKGROUND: Multiparametric magnetic resonance imaging (mpMRI) and targeted biopsies (TBs) facilitate accurate detection of significant prostate cancer (sPC). However, it remains unclear how many cores should be applied per target. OBJECTIVE: To assess sPC detection rates of two different target-dependent magnetic resonance imaging (MRI)/transrectal ultrasonography (TRUS)-fusion biopsy approaches (TB and target saturation [TS]) compared with extended systematic biopsies (SBs). DESIGN, SETTING, AND PARTICIPANTS: Retrospective single-centre outcome of transperineal MRI/TRUS-fusion biopsies of 213 men was evaluated. All men underwent TB with a median of four cores per MRI lesion, followed by a median of 24 SBs, performed by experienced urologists. Cancer and sPC (International Society of Urological Pathology grade group ≥2) detection rates were analysed. TB was compared with SB and TS, with nine cores per target, calculated by the Ginsburg scheme and using individual cores of the lesion and its "penumbra". OUTCOME MEASUREMENTS AND STATISTICAL ANALYSIS: Cancer detection rates were calculated for TS, TB, and SB at both lesion and patient level. Combination of SB + TB served as a reference. Statistical differences in prostate cancer (PC) detection between groups were calculated using McNemar's tests with confidence intervals. RESULTS AND LIMITATIONS: TS detected 99% of 134 sPC lesions, which was significantly higher than the detection by TB (87%, p = 0.001) and SB (82%, p < 0.001). SB detected significantly more of the 72 low-risk PC lesions than TB (99% vs 68%, p < 0.001) and 10% (p = 0.15) more than that detected by TS. At a per-patient level, 99% of men harbouring sPC were detected by TS. This was significantly higher than that by TB and SB (89%, p = 0.03 and 81%, p = 0.001, respectively). Limitations include limited generalisability, as a transperineal biopsy route was used. CONCLUSIONS: TS detected significantly more cases of sPC than TB and extended SB. Given that both 99% of sPC lesions and men harbouring sPC were identified by TS, the results suggest that this approach allows to omit SB cores without compromising sPC detection. PATIENT SUMMARY: Target saturation of magnetic resonance imaging-suspicious prostate lesions provides excellent cancer detection and finds fewer low-risk tumours than the current gold standard combination of targeted and systematic biopsies.

Simulated clinical deployment of fully automatic deep learning for clinical prostate MRI assessment.

OBJECTIVES: To simulate clinical deployment, evaluate performance, and establish quality assurance of a deep learning algorithm (U-Net) for detection, localization, and segmentation of clinically significant prostate cancer (sPC), ISUP grade group ≥ 2, using bi-parametric MRI. METHODS: In 2017, 284 consecutive men in active surveillance, biopsy-naïve or pre-biopsied, received targeted and extended systematic MRI/transrectal US-fusion biopsy, after examination on a single MRI scanner (3 T). A prospective adjustment scheme was evaluated comparing the performance of the Prostate Imaging Reporting and Data System (PI-RADS) and U-Net using sensitivity, specificity, predictive values, and the Dice coefficient. RESULTS: In the 259 eligible men (median 64 [IQR 61-72] years), PI-RADS had a sensitivity of 98% [106/108]/84% [91/108] with a specificity of 17% [25/151]/58% [88/151], for thresholds at ≥ 3/≥ 4 respectively. U-Net using dynamic threshold adjustment had a sensitivity of 99% [107/108]/83% [90/108] (p > 0.99/> 0.99) with a specificity of 24% [36/151]/55% [83/151] (p > 0.99/> 0.99) for probability thresholds d3 and d4 emulating PI-RADS ≥ 3 and ≥ 4 decisions respectively, not statistically different from PI-RADS. Co-occurrence of a radiological PI-RADS ≥ 4 examination and U-Net ≥ d3 assessment significantly improved the positive predictive value from 59 to 63% (p = 0.03), on a per-patient basis. CONCLUSIONS: U-Net has similar performance to PI-RADS in simulated continued clinical use. Regular quality assurance should be implemented to ensure desired performance. KEY POINTS: • U-Net maintained similar diagnostic performance compared to radiological assessment of PI-RADS ≥ 4 when applied in a simulated clinical deployment. • Application of our proposed prospective dynamic calibration method successfully adjusted U-Net performance within acceptable limits of the PI-RADS reference over time, while not being limited to PI-RADS as a reference. • Simultaneous detection by U-Net and radiological assessment significantly improved the positive predictive value on a per-patient and per-lesion basis, while the negative predictive value remained unchanged.