Validation, implementation, and impact of an AI model in routine practice for pathologic diagnosis of prostate cancer in an academic medical center

Abstract

Background

Pathological evaluation of prostate needle biopsies is labor-intensive and often requires ancillary immunohistochemistry (IHC), increasing cost and diagnostic turnaround time (TAT). Artificial intelligence (AI)-based decision-support tools may improve efficiency, but clinical deployment requires institutional validation and assessment of real-world impact.

Methods

Within a fully digital pathology practice, we performed institutional validation of an AI-assisted prostate biopsy decision-support tool using routine clinical cases, with pathologist-rendered diagnoses as ground truth. Following validation, the tool was implemented into the routine sign-out. A retrospective pre–post analysis compared prostate biopsy cases signed out during 3-month periods before and after implementation, excluding a transition month. Diagnostic TAT was defined as the interval from whole-slide image scan completion to final sign-out. IHC utilization was recorded. Weighted median TATs and IHC use were compared using standard statistical methods.

Results

The validation cohort met all predefined acceptance criteria, demonstrating high AI performance (sensitivity 91–100%, specificity 99%, positive-predictive value 98%, negative-predictive value 96%, area under the curve 0.97). Following clinical implementation, diagnostic TAT decreased by 30% and IHC utilization decreased by 38%.

Conclusions

Institutional validation and clinical implementation of an AI-assisted prostate biopsy decision-support tool were associated with significant reductions in diagnostic TAT and IHC utilization. When deployed as an adjunct within a digital workflow, AI assistance may enhance efficiency while preserving pathologist responsibility for final diagnosis.

Highlights

• •Institutional validation demonstrated high AI accuracy for prostate biopsy diagnosis.
• •Validation performance met acceptance criteria (sensitivity, specificity, AUC 0.97).
• •AI decision-support was integrated into routine digital pathology sign-out.
• •Diagnostic turnaround time decreased by 30% following AI implementation.
• •Immunohistochemistry utilization decreased by 38% after implementation.

Introduction

Prostate cancer is the second most common malignancy among men and a leading cause of cancer-related mortality. In the United States, more than 33,000 men died of prostate cancer in 2022, with mortality projected to increase to 36,500 by 2025.¹ Accurate and timely histopathological diagnosis is essential for appropriate risk stratification and therapeutic planning. Although prostate cancer often follows an indolent course, delays between biopsy, diagnosis, and treatment are common, and reducing time to diagnosis may facilitate earlier clinical decision-making and initiation of therapy.² Several artificial intelligence (AI)-based tools have been developed to improve diagnostic accuracy and efficiency in surgical pathology, many of which were developed specifically for prostate cancer diagnosis and detection.3, 4, 5 In the context of increasing case volumes and workforce constraints, AI offers the potential to enhance efficiency by assisting pathologists in slide review. AI models can systematically analyze whole-slide images (WSIs), highlight regions of interest that may be subtle or easily overlooked, and function as a “second set of eyes” to support diagnostic confidence. Early identification of suspicious or concerning areas has the potential to reduce reliance on confirmatory immunohistochemistry (IHC), particularly in cases where morphological findings are ultimately benign or unequivocal, thereby conserving resources and minimizing delays. Shorter diagnostic turnaround times (TATs) may lessen patient anxiety and support more timely treatment planning. Advances in AI also raise the possibility of future integration with molecular and quantitative pathology data to further enhance diagnostic accuracy and clinical utility.

Our Academic Department of Pathology has been an early adopter of digital pathology, having validated primary diagnosis by review of WSIs in Sept 2021 and transitioned to a fully digital workflow in July 2023. This digital infrastructure enabled institutional validation and implementation of an AI-based prostate biopsy decision-support system designed to generate probabilistic assessments of malignancy on WSIs. It has been emphasized that even commercially marketed pretrained AI tools with regulatory clearance for clinical application require rigorous local validation and careful workflow integration and should function as decision-support systems rather than autonomous diagnostic entities.6, 7

Here, we describe the local institutional validation, clinical implementation, and immediate workflow impact of an AI-assisted prostate biopsy cancer detection tool in routine practice at our academic medical center. Specifically, we report: (1) the results of an internal validation study comparing AI-generated probabilistic assessments with pathologist-rendered “ground-truth” diagnoses, (2) integration of the AI tool into an established digital pathology workflow, and (3) the impact of AI-assisted review on diagnostic TAT and IHC utilization immediately following clinical implementation.

Materials and methods

The diagnostic validity and clinical utility of the Ibex Prostate system have been previously established through blinded clinical validation studies and real-world clinical deployments conducted by the manufacturer and independent investigators.5, 8, 9 Based on the FDA CDRH 510(k) K241232 Decision Summary for Ibex Medical Analytics (January 24, 2025), the Galen™ Second Read™ (also referred to as Ibex Prostate Detect) was cleared as a software-only device to assist pathologists in analyzing prostate biopsies.5, 9, 10 Accordingly, the present study focused on local institutional verification and validation, consistent with lab requirements for deployment of externally validated AI-based diagnostic tools.

Before clinical deployment, the AI-assisted prostate biopsy tool (Ibex Prostate, Ibex Medical Analytics) underwent institutional verification and validation in accordance with lab policy and clinical lab standards at our institution. The purpose of this validation was to assess the performance of the AI model in generating probabilistic assessments of malignancy on digitized prostate biopsy slides within the local technical environment, workflow, and case mix.

Prostate needle biopsy cases received by the Surgical Pathology service over an approximately 2-month period were retrospectively selected for validation. All cases were digitized using Philips Ultra Fast Scanners and viewed in the Philips Image Management System. A minimum of 25 cases comprising at least 250 individual biopsies were included. Each case had been initially signed out by the assigned urologic surgical pathologist, viewing the WSI without the use of AI assistance. Following sign-out, all diagnoses were independently confirmed by a second pathologist (the senior author of this article), who was not involved in rendering the original diagnosis. This was designated the “ground-truth” diagnosis. For each biopsy “part” (representing the contents of a biopsy container), the ground truth was recorded and dichotomized as cancer present or cancer absent. Diagnoses categorized as atypical or suspicious were conservatively designated as non-cancer for validation purposes. WSI of the validation cases was securely transmitted to the Ibex Prostate platform via an encrypted virtual private network tunnel for analysis. The AI model was run on each slide and generated a probabilistic assessment of malignancy, assigning each biopsy part a classification of high, medium, or low likelihood of cancer based on predefined thresholds. When multiple slides were associated with a single biopsy part, the highest likelihood category assigned to any slide was used as the composite AI output for that part.

AI-generated likelihood assessments were compared with the pathologist-rendered ground-truth diagnosis at the biopsy-part level. Diagnostic performance was evaluated using sensitivity, specificity, positive-predictive value (PPV), and negative-predictive value (NPV), with the high-likelihood category used as the primary threshold for malignancy. Additional analyses assessed sensitivity when high- and medium-likelihood categories were combined and the NPV of the low-likelihood category for identifying biopsy parts without malignancy. Our institutional acceptance criteria required performance metrics of at least 0.85 sensitivity and specificity for the high-likelihood category, and at least 0.90 for combined high/medium sensitivity and low-likelihood NPV. Successful completion of these criteria supported clinical implementation of the AI tool as a decision-support system within the routine prostate biopsy workflow. Furthermore, a receiver operating characteristic (ROC) analysis was performed to evaluate the diagnostic performance of the AI system across ordinal confidence levels. AI interpretations were encoded as an ordered variable (low = 0, medium = 1, and high = 2). Sensitivity (true-positive rate) and false-positive rate were calculated at prespecified decision thresholds corresponding to these likelihood categories. ROC curves were constructed by plotting the true-positive rate against the false-positive rate across thresholds, and the area under the curve (AUC) was estimated using the trapezoidal rule. Ground-truth diagnosis served as the reference standard for all analyses. In addition to the ROC, the relationship between the AI confidence categories (low, medium, and high) and ground-truth diagnosis (malignant vs benign) was evaluated using the Cochran–Armitage trend test to specifically assess whether the probability of malignancy increased monotonically across ordered AI confidence levels, independent of threshold-based performance metrics. A two-sided p-value <0.05 was considered statistically significant.

Following successful validation, the AI-assisted prostate biopsy tool was integrated into routine clinical practice within the Surgical Pathology service. The department had previously transitioned to a fully digital workflow for primary diagnosis, in which all histological review and case sign-out were performed using WSI. This existing digital infrastructure enabled implementation of the AI tool without alteration of specimen processing, slide scanning, or reporting workflows. The AI system was made available to pathologists as an optional decision-support tool. All prostate biopsies received in surgical pathology were assigned to a GU pathologist as well as a nominal user called “Ibex-Prostate.” This assignment flags a case to be retrieved by the Ibex platform specifically for analysis and provides a probabilistic assessment of malignancy and visual heatmap overlays highlighting regions of interest on WSIs (Fig. 1).

Fig. 1.

Study design and workflow.

Schematic overview of the study, including institutional validation of the AI-assisted prostate biopsy decision-support tool, clinical implementation within a fully digital pathology workflow, and retrospective pre–post analysis to assess post-implementation impact on diagnostic TAT and immunohistochemistry utilization.

Use of the AI output was discretionary, and final diagnoses continued to be rendered exclusively by pathologists based on morphological assessment and clinical correlation.

To evaluate the immediate operational impact of AI-assisted review, a retrospective pre–post implementation study was performed comparing diagnostic TAT and IHC utilization for prostate needle biopsy cases. Cases signed out during a 3-month period immediately before AI availability were designated as the pre-AI cohort, whereas cases signed out during a 3-month period following implementation were designated as the post-AI cohort. All cases in both cohorts were interpreted using digital WSI, with AI assistance available only in the post-AI period.

Diagnostic TAT was defined as the interval between completion of whole-slide scanning and final pathologist sign-out. This definition was selected to isolate pathologist-controlled components of workflow and to minimize variability related to specimen accessioning, processing, or slide preparation. TAT was recorded for each case, and weighted median TATs were calculated for the pre- and post-AI periods. IHC utilization was assessed by recording the number of IHC stains ordered per case during each study period. Aggregate IHC utilization and average stains per case were compared between cohorts to evaluate changes in ancillary test ordering following AI implementation.

Statistical analyses were performed to compare diagnostic TAT and IHC utilization between the pre- and post-AI implementation periods. TAT, a continuous variable with a non-normal distribution, was summarized using medians and interquartile ranges and compared using the Mann–Whitney U test. IHC utilization was treated as a categorical variable and compared between groups using Fisher's exact test. All statistical tests were two-sided, with a significance threshold of p < 0.05, and exact p values are reported. Analyses were conducted as part of a retrospective quality and workflow assessment in accordance with institutional review board guidance.

Results

Validation results

A total of 318 biopsy parts derived from routine prostate needle biopsy cases were included in the institutional validation. Each part had three hematoxylin and eosin (H&E)-stained slides, which were all evaluated by the AI model. The AI model generated probabilistic assessments of malignancy for each biopsy slide, categorizing outputs as high, medium, or low likelihood of cancer. The final interpretation of the part was based on the highest category given to any of the three H&E slides for that part. For example, if the AI categorization for the three slides were, “low likelihood of cancer,” “high likelihood of cancer,” and “low likelihood of cancer,” respectively, then the entire part was categorized as, “high likelihood of cancer.” This method of adjudication converts a slide-level diagnosis into a part-level diagnosis and recapitulates what occurs in normal practice—for example, if slides level A1 and A9 are benign but slide A5 has cancer, the diagnosis of the entire part A is reported as cancer.

Of the 318 parts analyzed, 104 biopsy parts were classified as malignant and 214 as non-malignant (benign) based on pathologist-rendered ground truth diagnoses. Among 97 biopsy parts assigned a high likelihood of malignancy, 95 corresponded to malignant ground-truth diagnoses, whereas 2 were non-malignant. Of the 124 biopsy parts assigned a medium likelihood, 9 were malignant and 115 were non-malignant. All 97 biopsy parts assigned a low likelihood of malignancy were confirmed to be non-malignant on ground-truth diagnosis.

Using the high-likelihood category as the primary threshold, the AI model demonstrated a sensitivity of 0.91, a specificity of 0.99, a PPV of 0.98, and a NPV of 0.96 for malignancy detection. When high- and medium-likelihood categories were combined, sensitivity for malignancy reached 1.00, indicating that all biopsy parts containing cancer were assigned at least a medium likelihood of malignancy by the AI system. Conversely, biopsy parts assigned a low likelihood of malignancy demonstrated a NPV of 1.00, with no malignant cases categorized in this group.

Overall, all predefined acceptance criteria for validation performance were met or exceeded, supporting clinical implementation of the AI tool as a decision-support system within routine prostate biopsy workflow. The results of the validation study are shown in Table 1.

Table 1.

Prostate AI system local institutional validation results.

Validation of prostate AI system
Original diagnosis “ground truth”	Number (%)	AI-generated likelihood of malignancy	Number (%)
Total parts	318 (100)	High	97 (30.5)
Malignant	104 (32.7)	Medium	124 (39.0)
Benign	214 (67.3)	Low	97 (30.5)
2 × 3 Contingency table
Original diagnosis “ground truth”
AI system output (Likelihood of malignancy)	Malignant	Benign	Total
High	95	2	97
Medium	9	115	124
Low	0	97	97
Total	104	214	318
Metric		Value (%)
Sensitivity (High likelihood only)		95/104 (91.0)
Specificity		212/214 (99.0)
PPV		95/97 (98.0)
NPV		212/221 (96.0)
Sensitivity (High + Medium likelihood combined)		104/104 (100)

The ROC analysis also demonstrated excellent discriminative performance of the AI system. The AUC was 0.97, indicating strong separation between malignant and benign cases across increasing AI confidence levels. At the highest confidence threshold (high confidence only), the sensitivity was 91.3% with a false-positive rate of 0.9%, whereas inclusion of both high and medium confidence interpretations achieved 100% sensitivity with an expected increase in false-positive rate. Consistent with the high diagnostic discrimination observed on ROC analysis, there was a strong monotonic increase in the proportion of malignant cases with increasing AI confidence level (Cochran–Armitage trend test, p < 0.001). Malignancy rates rose from 0% in the low-confidence category to 7.3% in the medium-confidence category and 97.9% in the high-confidence category (Fig. 2), demonstrating effective ordinal risk stratification by the AI system.

Fig. 2.

AI validation performance.

Diagnostic performance of the AI-assisted prostate biopsy tool during institutional validation. Metrics include sensitivity (91–100%), specificity (99%), positive-predictive value (98%), negative-predictive value (96%), and area under the receiver operating characteristic curve (AUC = 0.97), demonstrating high accuracy across predefined acceptance criteria.

Post-AI implementation impact results

A total of 139 prostate needle biopsy cases were included in the immediate impact analysis, comprising 80 cases in the pre-AI period and 59 cases in the post-AI period. The study spanned 7 months, including a 3-month pre-implementation interval and a 3-month post-implementation interval, with a 1-month transition period excluded to minimize confounding related to user familiarization. All cases were reviewed using digital WSIs, with AI-assisted review available only during the post-implementation period.

Implementation of AI-assisted review was associated with a significant reduction in diagnostic TAT. The weighted median TAT decreased from 36.3 to 25.3 after implementation (−11.0 h; −30.3%; p = 0.0186) (Table 2, Fig. 3).

Table 2.

Turnaround time (TAT) before and after AI implementation.

Period	Months included	Cases (n)	Median TAT (hours)	Weighted median TAT (hours)	Absolute change (hours)	% change
Pre-AI	Aug–Oct 2024	80	30.2	36.3	−11.0	−30.3%
Post-AI	Dec 2024–Feb 2025	59	25.0	25.3	−11.0	−30.3%

Fig. 3.

Diagnostic turnaround time before and after AI implementation.

Comparison of diagnostic turnaround time (TAT) for prostate biopsy cases before and after AI implementation. Median TAT decreased by approximately 30% following deployment of AI decision-support within routine practice. TAT was defined as the interval from whole-slide image scan completion to final pathologist sign-out.

This difference was statistically significant (p = 0.0186). Reductions in TAT were observed consistently across the post-implementation months.

IHC utilization also declined following AI implementation. The average number of IHC stains ordered per case decreased from 2.37 in the pre-AI period to 1.47 in the post-AI period, corresponding to a 38% reduction in IHC utilization per case (Fig. 4).

Fig. 4.

Immunohistochemistry utilization pre- and post-implementation.

Immunohistochemistry utilization rates in prostate biopsy cases before and after AI implementation. Post-implementation analysis demonstrated a 38% reduction in IHC use, indicating improved diagnostic efficiency and resource utilization following AI-assisted workflow integration.

In addition to aggregate performance metrics, examination of individual validation cases provided important insight into how AI-generated probabilistic assessments should be interpreted in practice. Two instances were observed in which benign acinar proliferations were assigned a high likelihood of malignancy by the AI model (Fig. 5A, C). Both foci had been originally considered to be suspicious enough to merit IHC and were only revealed to be benign upon p63 staining of the basal cells of the glands (Fig. 5B, D). Conversely, among biopsy parts assigned a medium likelihood of malignancy, a small subset contained cancer, all of which were low grade and most of which were of small size (Fig. 6A–D). One larger focus in this category was a deceptively bland pseudo-hyperplastic variant type of adenocarcinoma, which the original pathologist had utilized immunostains to confirm as malignant (Fig. 6E, F).

Fig. 5.

Benign foci marked by the AI system with “high” probability of cancer.

Representative biopsy sections (A, C) show benign glands highlighted by the AI model as high probability for cancer (blue arrows). Corresponding immunohistochemical staining for p63 (B, E) demonstrates intact basal cell layers, confirming benign histology. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 6.

Malignant foci marked by the AI-algorithm as a “medium” probability of cancer.

Representative biopsy sections (A–D) in which malignant foci were assigned a medium probability score by the AI algorithm (blue arrows). These were all small foci of low-grade cancer. One focus (E) was a larger focus of an unusual variant (pseudo-hyperplastic) type of cancer. Immunohistochemical staining for p63 (F) shows absence of basal cells, confirming carcinoma. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Discussion

The last 7 years have seen increasing availability of AI models to assist in pattern recognition tasks in surgical pathology.¹¹ It is not accidental that the detection of prostate cancer is one of the earliest clinical applications of this technology.⁴ The traditional method of evaluation of these biopsies is particularly tedious and requires a high degree of vigilance on the part of the pathologist because the microscopic appearance of prostate cancer can be notoriously subtle, the foci of cancer can be quite small, and the consequence of missing cancer can be serious. These are exactly the types of tasks where AI excels, performing a thorough assessment of each patch and pixel without fatigue. These qualities are clearly beneficial in a variety of other situations in diagnostic pathology, hence the growing number of AI models to assist in the detection of other types of cancer and quantification of biomarkers.12, 13 In fact, it is the availability of AI tools and a recognition of their power and inevitability that has stimulated a rapidly growing interest in the adoption of digital pathology.

Given the ubiquitous and somewhat standardized nature of the deep learning methods by which these AI models are trained, the relative strength of one model over another rests in the quality, size and diversity of the underlying training data. The Ibex Prostate model described in this study was developed with a large, multi-institutional dataset of WSI and has been subjected to technical and clinical validation in a variety of settings across the globe.

Regulatory clearance by the FDA provides a strong measure of confidence in the AI model, and it can therefore be considered “clinical grade.” An important and helpful feature of this software is good integration with existing workflow and a visually intuitive heatmap presentation of the outcome. Finally, the additional evaluations performed by the model—reporting not only the presence or absence of cancer but also the subtype of cancer, amount, grade, and detection of perineural invasion—these features provide valuable ancillary information that facilitates a complete evaluation of the biopsy. It is important to note, however, that this additional ancillary information is informational only and was not submitted for clearance, and therefore has not been evaluated by the FDA.

A key strength of this study is the structured institutional validation performed before clinical deployment. The AI system functioned exclusively as a decision-support tool, providing probabilistic assessments of malignancy, and did not render diagnoses; all final diagnoses were made by pathologists. Local validation of AI-assisted tools is a critical step in mitigating patient safety risk and ensuring acceptable performance within the local technical environment, case mix, and workflow. Validation was conducted using routine clinical prostate biopsy material interpreted independently by pathologists, with conservative ground-truth definitions and predefined acceptance thresholds for sensitivity, specificity, and predictive values. This approach aligns with recommended best practices for clinical AI validation, which emphasize local performance assessment rather than reliance on vendor-reported metrics alone.⁷ By confirming that the AI model consistently generated reliable probabilistic assessments of malignancy and met predefined acceptance criteria, validation supported its use as a decision-support system while preserving pathologist responsibility for final diagnosis. Importantly, our validation strategy reflects emerging consensus that AI tools intended for clinical use should be evaluated under real-world conditions that mirror intended use, including local scanning platforms, image management systems, and interpretive workflows. Professional organizations and regulatory bodies have emphasized that validation of AI in pathology should be analogous to validation of other lab tools, with clear documentation, predefined performance criteria, and ongoing quality assurance. The ROC findings indicate that AI confidence scores provide robust discriminatory information that can be flexibly applied to clinical workflows. High-confidence AI predictions demonstrated high specificity with minimal false positives, supporting their use for streamlined diagnostic review or prioritization. Lowering the decision threshold increased sensitivity at the expense of specificity, allowing tailoring of AI integration based on clinical context and risk tolerance. Overall, the high AUC underscores the reliability of AI confidence stratification as a clinically meaningful adjunct to pathologist interpretation. Our findings are consistent with those reported by Kris Lami and colleagues⁶ in 2024, who evaluated the performance of the Galen Prostate algorithm in a retrospective Japanese cohort and reported an AUC of 0.969 for prostate cancer detection. In that study, the AI system demonstrated strong discriminatory capability for identifying malignant biopsy cores and additionally detected clinically relevant pathological features such as perineural and lymphovascular invasion. Like our results, their findings support the concept that AI-derived probability or confidence scores can reliably distinguish malignant from benign tissue and may serve as a practical decision-support tool to assist pathologists during routine diagnostic review. The concordance between our AUC results and those reported in independent validation cohorts across different geographic populations and clinical settings further supports the robustness and generalizability of this AI model. The validation process described here provided a foundation for safe implementation and contextualizes the observed post-implementation workflow improvements, reinforcing that efficiency gains were achieved within a framework of verified performance rather than untested adoption.

The exceptional cases described in Fig. 5, Fig. 6 serve as reminders that AI outputs represent probabilistic assessments rather than definitive determinations and reinforce the necessity of pathologist oversight. In our validation cohort, only two biopsy parts assigned a high likelihood of malignancy were originally classified as benign, resulting in a high PPV while illustrating the expected behavior of a sensitive screening-oriented tool. Similarly, the examples of cancer illustrated in Fig. 6 highlight the intermediate nature of the medium-likelihood category and its intended role as a prompt for closer morphological scrutiny rather than a binary classification. Importantly, all biopsy parts assigned a low likelihood of malignancy were confirmed to be benign on final pathologist interpretation, supporting the use of the low-likelihood output as a more reliable indicator for the absence of malignancy within the context of expert review.

Taken together, these case-level observations complement the overall validation metrics, which demonstrated high sensitivity, specificity, and predictive values. When high- and medium-likelihood categories were considered together, sensitivity for malignancy reached 1.00, indicating that all biopsy parts containing cancer were flagged by the AI system at a level warranting pathologist attention. These findings supported the use of the AI tool as a screening and prioritization aid rather than a diagnostic arbiter and informed its integration into clinical workflow.

Following validation, the AI tool was incorporated into routine prostate biopsy evaluation such that AI-generated probabilistic assessments and heatmap visualizations could be reviewed alongside WSIs before ordering immunohistochemical stains. This approach allowed pathologists to incorporate AI output early in the interpretive process while retaining full discretion over its use. The tool also provided access to additional features, including Gleason pattern visualization and identification of perineural invasion, further supporting efficient review without supplanting morphological assessment. To our knowledge, this represents one of the earliest implementations of such a workflow for routine prostate biopsy diagnosis by an academic pathology department in the USA. Importantly, the use of the AI tool remains entirely discretionary, allowing pathologists to determine the extent to which AI outputs inform their review on a case-by-case basis. This flexible integration model reflects real-world practice and mitigates the risk of over-reliance on automated outputs. Within this framework, post-implementation analysis demonstrated reductions in TAT and IHC utilization, suggesting that early access to AI-generated probabilistic assessments may enhance diagnostic efficiency while preserving pathologist judgment and responsibility.

Implementation of the prostate AI decision-support tool was associated with a significant reduction in aggregate diagnostic TAT for prostate biopsies in routine clinical practice. This finding is consistent with prior real-world implementation studies, including the work by Deman et al.,14, 15 which demonstrated improvements in TAT following an initial period of user familiarization with AI-assisted tools.

In our study, the weighted median TAT decreased from 36.2 to 25.4 h, representing an absolute reduction of 10.9 h, or approximately 30%. This improvement was statistically significant (p < 0.05).

The AI model (Ibex Prostate) described in this study has performance characteristics that are very similar to the AI model (Paige Prostate) that was first approved by the FDA.4, 16, 17 Although we did not evaluate or validate that model in our institution, our experience and results are very similar to those reported previously using Paige Prostate.

In the present study, the average number of IHC stains per case decreased from 2.37 to 1.47 following implementation, representing a 38% reduction in IHC utilization. This finding is consistent with prior studies evaluating AI-assisted prostate biopsy review. Eloy et al. and Flach et al. similarly reported statistically significant reductions in IHC use for both malignant and non-malignant cases when AI decision-support was incorporated into the diagnostic workflow.18, 19

The observed improvements in TAT and ancillary test utilization likely reflect multiple contributing factors rather than a single mechanism. AI-generated probabilistic assessments and heatmap visualizations may facilitate more efficient identification of regions of interest on WSIs, particularly in cases with limited tumor volume. Additionally, increased diagnostic confidence in biopsy parts assigned a low likelihood of malignancy may reduce time spent reviewing non-diagnostic areas and decrease reliance on confirmatory IHC. Importantly, the AI tool functioned exclusively as a decision-support system; all final diagnoses were rendered by pathologists integrating morphological findings, AI outputs, and clinical context.

Reduced IHC utilization has implications beyond TAT alone. Decreased reliance on ancillary stains may lower material and labor costs, reduce strain on histology workflows, and improve overall lab efficiency. Furthermore, AI-assisted visualization of suspicious and non-suspicious regions may have educational value, particularly for less-experienced pathologists, by reinforcing morphological patterns associated with malignancy and benign mimickers. Whereas such educational effects were not formally assessed in this study, they represent an important area for future investigation.

Our findings add to a growing body of literature supporting the role of AI-assistance tools in enhancing efficiency across a range of histopathological applications. Multiple studies have demonstrated improvements in TAT and workflow efficiency following AI integration in both academic and non-academic settings for various histopathological entities.14, 16, 20, 21, 22, 23, 24 Collectively, these observations suggest that AI decision-support may be particularly beneficial in high-volume practices and resource-constrained environments, although the extent of benefit likely varies by case mix, workflow, and user experience.

This study has several limitations that should be considered when interpreting the findings. First, the analysis was conducted at a single academic medical center operating within a fully digital pathology workflow. Whereas this environment facilitates efficient AI integration, the results may not be directly generalizable to institutions with different levels of digital maturity, case mix, or staffing models. Multi-institutional studies will be necessary to confirm reproducibility across diverse practice settings.

Second, the post-implementation impact analysis was limited to a relatively short time frame and a moderate number of cases. Although statistically significant reductions in diagnostic TAT and IHC utilization were observed, longer-term studies are needed to evaluate the durability of these effects, potential learning curves, and adaptation over time. In addition, downstream clinical outcomes, including time to treatment initiation and patient-reported measures, were not assessed and remain outside the scope of this study.

Third, TAT was defined as the interval between completion of WSI scanning and final pathologist sign-out. This definition was intentionally selected to isolate workflow components under pathologist control; however, it does not account for upstream variables such as specimen accessioning, tissue processing, or slide preparation, which are unaffected by AI assistance. Consequently, the observed reductions reflect improvements in a defined segment of the diagnostic pathway rather than the entirety of lab operations.

Finally, this study evaluated a single commercially available AI platform validated and implemented for a specific clinical use case. Prior studies have demonstrated that AI tools may support a range of functions, including pre-screening, case prioritization, and anticipatory ordering of ancillary studies, highlighting opportunities for broader workflow optimization.25, 26, 27

AI systems vary substantially in design, purpose, and implementation strategy, and generalizability across platforms must be evaluated cautiously. Some institutions employ commercially available tools, whereas others utilize in-house models, such as those described by Mayall et al.²⁶ Accordingly, our findings are specific to the Ibex Prostate platform and should not be extrapolated to other AI systems without independent validation. Importantly, the AI tool functioned exclusively as a decision-support system, and all final diagnoses were rendered by pathologists.14, 20 As with any AI-assisted workflow, continued oversight, periodic revalidation, and performance monitoring remain essential to ensure safe and effective clinical use.

Conclusion

In this study, we describe the institutional validation, clinical implementation, and immediate post-implementation impact of an AI-assisted decision-support tool for prostate biopsy evaluation in a fully digital academic pathology practice. Through a structured validation process, the AI model demonstrated reliable performance in generating probabilistic assessments of malignancy, supporting its safe deployment as an adjunct to pathologist interpretation.

Following implementation, AI-assisted review was associated with significant reductions in diagnostic TAT and IHC utilization, suggesting improved workflow efficiency and resource optimization. These improvements were observed without altering the central role of the pathologist, who remained solely responsible for rendering final diagnoses.

Taken together, these findings support the responsible integration of AI as a decision-support tool in routine prostate biopsy practice. When preceded by appropriate institutional validation and implemented within a digital pathology framework, AI assistance has the potential to enhance efficiency while preserving diagnostic quality and professional oversight. Further multi-institutional studies and longer-term evaluations are warranted to define the broader and sustained impact of AI-assisted pathology workflows.

Ethical approval

This validation and implementation study was performed as part of an institutional quality improvement and quality assurance initiative using retrospectively collected, de-identified data. The Institutional Review Board reviewed the project and determined that it did not meet the regulatory definition of human subjects' research under 45 CFR 46.102. Accordingly, IRB oversight and informed consent were not required.

Funding

No external funding was obtained for this study.

Use of generative artificial intelligence

The authors used generative AI tools (ChatGPT, OpenAI) to assist with language refinement, formatting, and editorial improvement of the manuscript. The AI tool was not used for data analysis, data interpretation, generation of results, image analysis, or statistical calculations. All scientific content, analyses, conclusions, and interpretations were independently developed, verified, and approved by the authors, who take full responsibility for the integrity and accuracy of the work.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this article.