Nurse-Validated Handheld POCUS Artificial Intelligence (AI) for Automated Prostate Volume Estimation and Hydronephrosis Detection

Masaru Matsumoto, PhD RN1,2*; Motofumi Suzuki, PhD MD3; Mikihiko Karube, M.Eng2; Hiroaki Gotoh, MPharm MT RMS4; Masaya Sato, PhD MD5; Tamaki Kobayashi, PhD MT RMS4; Minoru Masuda, MHS RN6,7; Sanae Wakita, RN6; Atsuo Kawamoto, PhD RT RMS8; Yuka Miura, PhD RN2,9; Nao Tamai, PhD RN2,10,11; Hiromi Sanada, PhD RN CWOCN7,11,12

(1) Department of Well-being Nursing, Graduate School of Nursing, Ishikawa Prefectural Nursing University, Kahoku, Ishikawa, Japan

(2) Former Department of Imaging Nursing Science, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

(3) Department of Urology, Mishima General Hospital, Mishima, Shizuoka, Japan

(4) Department of Clinical Laboratory Medicine, The University of Tokyo Hospital, Tokyo, Japan

(5) Department of Clinical Laboratory Medicine, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

(6) Former Department of Nursing, The University of Tokyo Hospital, Tokyo, Japan

(7) Department of Gerontological Nursing/Wound Care Management, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

(8) Division of Ultrasound and Department of Diagnostic Imaging, Tokyo Medical University Hospital, Tokyo, Japan

(9) Research Center for Implementation Nursing Science Initiative, Fujita Health University, Toyoake, Aichi, Japan

(10) Department of Nursing, Graduate School of Medicine, Yokohama City University, Yokohama, Kanagawa, Japan

(11) Former Global Nursing Research Center, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

(12) Ishikawa Prefectural Nursing University, Kahoku, Ishikawa, Japan

*Corresponding Author: Dr. Masaru Matsumoto (email: matumoto@ishikawa-nu.ac.jp)


Download article PDF – POCUS Journal 2026;11(2):60-70

DOI: https://doi.org/10.24908/pocusj.v11i02.20395


Abstract

Background: Artificial intelligence (AI) can detect hydronephrosis and estimate prostate volume on ultrasound, but most systems rely on expert-acquired images from conventional console ultrasound systems. We developed an on-device AI model for handheld point-of-care ultrasound (POCUS) and prospectively evaluated it on nurse-acquired scans for bedside triage/decision support. Methods: The model was trained using retrospective console images (101 renal and 100 prostate cases) and prospective POCUS development clips (44 renal and 88 prostate patients), then evaluated prospectively on ward nurse-acquired renal and suprapubic POCUS videos after brief instruction. Reference standards were blinded expert adjudication of hydronephrosis and expert planimetry for prostate volume. Primary outcomes were sensitivity/specificity for hydronephrosis detection and screening performance for severe prostate enlargement (≥50 mL). Results: For hydronephrosis detection, 32 evaluable patients were analyzed (one kidney per patient; prevalence 46.9%). The AI achieved a sensitivity of 93.3% (95% CI 70.2–98.8) and a specificity of 82.4% (59.0–93.8). For severe prostate enlargement (≥50 mL; n = 28), sensitivity and specificity were both 71.4%. Agreement for continuous prostate volume was moderate: ICC(A,1) = 0.789; Bland–Altman bias +5.1 mL, with 95% limits −25.0 to +35.1 mL. Conclusions: On nurse-acquired handheld POCUS, AI showed high sensitivity for hydronephrosis and moderate agreement for prostate-volume estimation, supporting its feasibility for triage support while requiring further validation.

Introduction

Lower urinary tract symptoms (LUTS) are common urological problems in older adults receiving home care and occur more frequently than in the general population [1]. In Japan, the need for urinary care in home settings is growing as healthcare increasingly shifts outside of hospitals. The proportion of older adults is projected to exceed 30% by 2025 and approach 40% by 2065 [2]. LUTS, encompassing storage, voiding, and post-void symptoms, can substantially impair quality of life [3,4]. Because management depends on symptom type and cause, accurate characterization is essential [5]. When bladder emptying fails and urine remains undrained for prolonged periods, complications such as acute retention, upper urinary tract dilatation and urinary tract infection may ensue [6,7]. In home care, early recognition by visiting nurses is critical for preventing disease progression.

POCUS has expanded nurses’ ability to assess LUTS at the bedside, especially with the development of handheld systems that are portable, inexpensive, and readily available. Current guidelines recommend measuring postvoid residual urine volume to screen for voiding dysfunction, typically using a threshold of approximately 100 mL [8,9]. Handheld devices can quantify small bladder volumes, including with automated measurement functions. Nursing education programs for handheld POCUS are being developed [10,11].

However, bladder-only imaging cannot determine the etiology or severity. It does not establish whether outlet obstruction from prostate enlargement is present, nor whether stasis has propagated upstream to the kidneys, causing hydronephrosis. Patients benefit from a multiorgan view that includes the bladder, prostate, and kidneys for clinical decision-making, triage, catheter management, and referral. An algorithmic approach using POCUS to measure bladder volume while also assessing prostatic hypertrophy and hydronephrosis has been proposed to support timely selection of voiding care in home care settings [10]. However, consistent image interpretation requires training and time.

Artificial intelligence (AI) has shown promise for refining and streamlining interpretation of urological health metrics, including hydronephrosis detection and automatic prostate-volume (PV) estimation. However, most prior studies used console ultrasound systems with expert-acquired images, focused on single-organ tasks, and provided limited evidence for generalizability to nurse-acquired handheld scans in real-world settings [11–13]. Although AI-based hydronephrosis detection and deep-learning frameworks for prostate-volume estimation have been reported, integrated prostate-kidney assessment on handheld platforms evaluated using nurse-acquired images remains largely unreported [13–17].

To address this gap, we developed an AI system purpose-built for handheld POCUS and prospectively evaluated it using nurse-acquired scans after brief instruction. Our objectives were to (i) determine diagnostic performance for kidney-level hydronephrosis detection and (ii) assess agreement with an expert reference for patient-level PV estimation, including a prespecified ≥50 mL triage flag, in routine bedside practice.

Methods

Study design and setting

We conducted a diagnostic accuracy study using a development–validation sequence at a national university hospital in Japan. Phase 1 consisted of image collection, annotation, and algorithm development, while Phase 2 was a prospective clinical evaluation of nurse-acquired handheld scans. Reporting follows key STARD items where applicable [18].

Participants and sampling (evaluation cohort)

During the evaluation phase, consecutive adult inpatients in the urology ward were prospectively enrolled between April and September 2021. Eligible patients had a clinical indication for bedside renal and/or prostate ultrasound and were able to undergo POCUS scanning. Patients who could not provide consent or whose scans were non-diagnostic because of inadequate acoustic windows were excluded. Ward nurses with limited prior ultrasound experience performed all scans after a 30-minute device-use lecture. A STARD-style flow diagram is shown in Figure 1.

Figure 1. STARD-style flow diagram for the prospective evaluation cohort. Kidney imaging: 35 patients consented; 3 were not scanned within the study window (time constraints), and 32 were included in analysis (indeterminate = 0). Prostate imaging: 36 patients consented; 3 were not scanned within the study window, 33 were scanned, 5 were excluded as indeterminate/non-diagnostic, and 28 were included in analysis.

Intended use and clinical role of the index test

The index test is an automated algorithm intended for handheld POCUS to support nurse-led bedside assessment. It provides two functions: hydronephrosis detection (present/absent), and prostate volume estimation with severity classification, labelling severe enlargement when volume is ≥50 mL. The ≥50 mL threshold was selected a priori to identify more pronounced prostate enlargement relevant to triage and referral decisions in nurse-led bedside assessment, rather than to define any prostate enlargement. We acknowledge that lower thresholds, such as 30 mL, are commonly used to define prostate enlargement in other clinical contexts; evaluation of lower thresholds was beyond the scope of this pilot study. The intended clinical roles are triage, decision support for catheter management, and referral in home and ward settings.

Image acquisition (evaluation cohort; index test data)

Nurses acquired handheld scans using a wireless POCUS device with a convex probe (iViz air, FUW-1; Fujifilm, Tokyo, Japan; 2–5 MHz) using the abdomen preset according to a standardized protocol. This device allows adjustment of gain and depth only; these settings were optimized by each operator at the bedside. The longitudinal and transverse sweeps of both kidneys were recorded as 800 × 600-pixel videos, sampled every 12 frames (approximately 0.5 s), and resized to 512 × 384 pixels for analysis. For this pilot patient-level analysis, one kidney per patient was evaluated to avoid within-patient correlation in a small cohort; the side with the clearest acoustic window was selected. We acknowledge below that this selection may introduce bias compared with bilateral screening. Kidneys with indeterminate views were excluded, as specified. For the prostate, suprapubic transverse and sagittal views were obtained with the bladder filled when feasible; however, bladder volume was not standardized or systematically recorded. Videos were captured at 800 × 600 pixels, frames were sampled every 24 frames (approximately 1.0 s) and resized to 512 × 384 pixels for analysis. Operator identity and prior experience were not recorded in this analysis.

Reference standards and blinding

For hydronephrosis, a certified sonographer who was blinded to all index test outputs and metadata reviewed the nurse-acquired clips. They determined the presence or absence per patient based on dilatation of the renal pelvis and calyces visualized on the handheld videos. Borderline renal pelvic fullness or mild pelviectasis was adjudicated according to the binary expert reference standard and was not modelled as a separate category. For prostate volume, the same blinded expert performed manual segmentation and planimetry on selected transverse and sagittal frames. The prostate volume was calculated using the ellipsoid formula (0.52 × width × height × length). Severe prostate enlargement was defined as ≥50 mL a priori. The time interval between the index and reference was zero because both were derived from off-line readings of the same acquisition. Rules for non-evaluable cases were prespecified as follows: kidneys for which hydronephrosis could not be adjudicated because of inadequate acoustic windows were classified as indeterminate and excluded from accuracy estimates. The number of indeterminate kidneys is reported in the results section. For prostate volume, frames with inadequate anatomical delineation were discarded. If no valid orthogonal frame pair could be identified, the case was considered indeterminate and excluded from accuracy analyses. The counts of such cases are reported in the results section.

Development datasets (training only; not used for accuracy estimates)

Separate datasets were assembled at the same institution for algorithm development. Retrospective high-end conventional console ultrasound still images acquired by certified technologists included 101 hydronephrosis-positive renal still-image cases and 100 prostate still-image cases (paired transverse and sagittal views). In addition, prospective handheld development clips were collected using the same handheld device as the evaluation cohort (iViz air) from consenting patients. The kidney handheld dataset comprised 44 patients with both kidneys scanned (88 clips, 10 s each; 20 hydronephrosis-positive and 68 negative). Thus, the retrospective console renal data and handheld renal data were different data units (still-image cases versus video clips sampled into frames) and should not be interpreted as 121 hydronephrosis-positive clips. The prostate handheld dataset comprised 88 patients (10 s suprapubic clips). For handheld development clips, still frames were extracted at 12 frames/s for annotation and training; the training sets used 3,341 kidney frames and 7,420 prostate frames. Development data were split at the case/patient level into training and validation sets (kidney: 128/17; prostate: 167/21). The prospective nurse-acquired evaluation cohort served as an independent test set (kidney: 32 patients; prostate: 28 patients). Details of dataset sources and case/patient-level splits are summarized in Table 1. These developmental data were used solely for training and tuning. All accuracy estimates in this study were derived from prospectively collected nurse-acquired handheld scans in the ward.

Table 1. Composition of development and evaluation datasets (Phase 1 and Phase 2). Phase 1 development sources: kidney—101 hydronephrosis-positive high-end conventional console still-image cases (retrospective) and 44 handheld cases with both kidneys scanned (88 video clips, 10 s each; 20 hydronephrosis-positive and 68 negative). These retrospective console data and handheld video data represent different data units (still-image cases versus video clips sampled into frames) and were not pooled as clips. Prostate—100 high-end conventional console still-image cases with paired transverse and sagittal views (retrospective) and 88 handheld cases (10-s suprapubic clips). Handheld clips were acquired with the same handheld device as the prospective evaluation cohort. Case/patient-level split (train/validation/independent prospective test): kidney 128/17/32; prostate 167/21/28. Approximate number of frames used for training: kidney 3,341; prostate 7,420. Phase 2 prospective evaluation: kidney cohort—35 consented/enrolled, 3 not scanned within the study window, 32 included in analysis (indeterminate/non-diagnostic, 0); prostate cohort—36 consented/enrolled, 3 not scanned within the study window, 5 indeterminate/non-diagnostic, and 28 included in analysis.
TaskSplitCases/patients (n)Dataset/acquisitionNotes
Hydronephrosis (kidney)Training128Phase 1 development (console still images + handheld clips)3,341 labelled frames used for training; handheld clips were converted to still frames at 12 frames/s.
Hydronephrosis (kidney)Validation17Phase 1 development (console still images + handheld clips)Used for tuning and pre-specification of operating points.
Hydronephrosis (kidney)Prospective evaluation (test)32Phase 2 nurse-acquired handheld clipsOne kidney per patient; indeterminate=0.
Prostate volume estimationTraining167Phase 1 development (console still images + handheld clips)7,420 labelled frames used for training; handheld clips were converted to still frames at 12 frames/s.
Prostate volume estimationValidation21Phase 1 development (console still images + handheld clips)Used for tuning and prespecification of operating points.
Prostate volume estimationProspective evaluation (test)28Phase 2 nurse-acquired handheld clipsExcluded indeterminate = 5 (non-diagnostic).

Annotation for development

A single certified sonographer blinded to the clinical information and index test outputs annotated the training data using LabelMe (open-source Python/Qt implementation by Kentaro Wada; exact software version was not retained in the study records). For handheld videos, frames were extracted at 12 frames per second to generate still images and representative frames were selected for polygon annotation. For kidneys, polygonal regions of interest delineated the entire kidney and the anechoic intrapelvicalyceal urine. Frames were sampled every 0.5 s. For the prostate, polygons outlined the prostate and bladder on both transverse and sagittal frames. Frames were sampled every 1.0 s. Annotations were exported as JSON and converted to pixel-wise binary masks at the native resolution for model training. Figure 2 shows representative examples of training data.

Figure 2. Training annotations for prostate and kidney on handheld ultrasound. (A) Transverse suprapubic view of the prostate. Polygonal overlays drawn in LabelMe delineate the prostate gland. The bladder lumen is additionally contoured to clarify the prostate–bladder interface during labelling. (B) Sagittal suprapubic view of the prostate using the same scheme, with separate polygons for prostate and bladder to maintain anatomical separation across planes. (C) Renal section. Polygons outline the entire kidney and the anechoic renal pelvic urine collection corresponding to hydronephrosis. All annotations were created as polygon regions of interest in LabelMe and exported as binary masks for supervised training of the segmentation model. Images are representative frames sampled from nurse-acquired handheld video clips according to the study protocol. Colored overlays are illustrative only. Model training used the underlying pixel-wise masks.

Algorithm and implementation

We implemented a frame-level semantic segmentation network comprising a U-Net decoder and a ResNeXt-50 encoder pretrained on natural image data [19]. Data augmentation consisted of random brightness, contrast, sharpness changes, horizontal flips, and random erasing. The inputs were resized to 512 × 384 pixels, while preserving the aspect ratio. Training used the Adam optimizer with an initial learning rate of 1 × 10–⁵ for 100 epochs, was implemented in Keras, and ran on a single NVIDIA GTX 1080 Ti GPU.

Post-processing and case-level aggregation

For hydronephrosis, the model produced a probability mask for intrarenal urine within each frame. A kidney was classified as positive if any frame within its clip exceeded a prespecified probability or area threshold determined on a held-out development set. Otherwise, it was classified as negative. The current implementation was designed for binary hydronephrosis detection and did not separately classify mild pelviectasis/renal pelvic fullness from clinically significant hydronephrosis. Prostate volume, width, anteroposterior height, and length were extracted from the segmented transverse and sagittal frames. The per-patient volume was computed using the ellipsoid formula. The per-patient volume was defined as the median value across valid frame pairs within a study. The ≥50 mL threshold for severe enlargement of the prostate and the frame-selection rule were fixed before clinical evaluation.

Outcomes

Primary outcomes at the prespecified operating points were sensitivity and specificity for hydronephrosis detection on a per-patient basis and sensitivity, specificity, and overall accuracy for severe prostate enlargement (≥50 mL) on a per-patient basis. The secondary outcome was agreement for continuous prostate volume (ICC(A,1), Bland–Altman, Pearson’s r).

Statistical analysis

Participant characteristics were summarized at the patient level. Diagnostic accuracy was calculated for each patient with hydronephrosis and each patient with prostate outcomes. Sensitivity, specificity, positive predictive value, negative predictive value, and 95% confidence intervals (CIs) were reported using the Wilson method. For continuous prostate volume, agreement between the index test and the reference was quantified using the intraclass correlation coefficient (ICC(A,1)), a two-way mixed-effects, absolute-agreement, single-measure model in which subjects were treated as random and the two raters (AI and the reference reader) as fixed. This reflected how closely a single AI measurement reproduced the reference on the same subject and penalized both constant bias and scale differences. The 95% CI for the ICC was obtained via percentile bootstrapping with 1,000 resamples. Correlations were summarized using Pearson’s r with 95% CIs derived from Fisher’s z transformation, and Bland–Altman analysis reported bias and 95% limits of agreement. Missing or indeterminate results were handled according to the prespecified rules described in the methods section. No formal a priori sample size calculation was performed. The evaluation was conceived as a pilot technical and implementation study to assess prespecified operating points chosen in the held-out development data. Therefore, the primary analysis reports fixed-threshold diagnostic accuracy rather than ROC-optimized thresholds; ROC analyses were not undertaken in this pilot report and should be incorporated into larger validation studies.

Results

Thirty-five patients were enrolled for kidney imaging. After consent was obtained, three patients could not be scanned within the study window due to time constraints and were excluded. Of the remaining 32 patients with complete baseline data, the mean age was 71.8 ± 10.7 years (median, 74.0; interquartile range (IQR), 69.8–79.0) and the mean body mass index (BMI) was 23.0 ± 3.3 kg/m² (median, 23.4; IQR, 21.1–24.7). In the primary diagnosis category, genitourinary conditions accounted for 18/32 (56.3%) cases, and neoplasms accounted for 14/32 (43.7%) cases. No kidneys were classified as indeterminate (non-diagnostic) in the evaluable cohort.

The reference prevalence of hydronephrosis was 46.9% (15/32 patients). At the prespecified operating point, the index test achieved sensitivity 93.3% (95% CI, 70.2–98.8), specificity 82.4% (95% CI, 59.0–93.8), PPV 82.4% (95% CI, 59.0–93.8), negative predictive value (NPV) 93.3% (95% CI, 70.2–98.8), and overall accuracy 87.5% (95% CI, 71.9–95.0). Representative qualitative examples are shown in Figure 3. The corresponding 2 × 2 contingency table is shown in Table 2.

Figure 3. Qualitative outcomes of AI hydronephrosis detection on handheld ultrasound. Panels are arranged by rows (case outcome) and columns (from left to right: AI mask, original ultrasound, ground truth). Row 1 (A–C) shows a True Positive kidney; Row 2 (D–F) a True Negative; Row 3 (G–I) a False Positive; Row 4 (J–L) a False Negative. In the AI panels (A, D, G, J), red indicates the segmented kidney and yellow indicates anechoic renal pelvic urine collection (dilatation). In the ground truth panels (C, F, I, L), the yellow polygon is the expert annotation used to derive reference labels. All examples are nurse-acquired handheld scans selected to illustrate typical success and failure modes, such as subtle dilation or poor acoustic windows. Colours are for visualization only. Classifications were computed from the underlying binary masks according to prespecified rules. AI, artificial intelligence.
Table 2. Hydronephrosis (per patient): 2 × 2 contingency table for the index test vs reference. Evaluable patients (one kidney per patient): N = 32 (TP = 14, FP = 3, FN = 1, TN = 14). The sensitivity, specificity, PPV, and NPV were reported with 95% confidence intervals (Wilson method).
Reference
+ (present)− (absent)
AI+ (present)143
− (absent)114

A total of 36 patients were enrolled for prostate imaging. After providing consent, three patients could not be scanned within the study window because of time constraints and were excluded. Nurse-acquired POCUS was obtained in 33 patients, of whom five (15.2% of scanned patients) were excluded due to complete or partial absence of the AI-estimated prostate area, non-detection of the bladder region, or frames unsuitable for reference planimetry.

Twenty-eight men were included in the prostate analysis. Mean age was 71.9 ± 8.2 years (median, 72.5; IQR, 67.2–78.0). Mean BMI was 22.8 ± 3.0 kg/m² (median, 22.4; IQR, 21.0–24.0). In the primary diagnosis category, neoplasms accounted for 16/28 (57.1%) patients, genitourinary conditions for 9/28 (32.1%), and other conditions for 3/28 (10.7%). Within the neoplasm category, eight patients were evaluated for suspected malignancy, and one had a testicular tumour.

In the per-patient prostate analysis, 28 paired measurements of the reference and AI-estimated volumes were available. Using the prespecified threshold of ≥50 mL to define severe prostate enlargement, the prevalence by reference was 25.0% (7/28). At this operating point, the AI achieved sensitivity 71.4% (95% CI, 35.9–91.8), specificity 71.4% (95% CI, 50.0–86.2), positive predictive value 45.5% (95% CI, 21.3–72.0), negative predictive value 88.2% (95% CI, 65.7–96.7), and overall accuracy 71.4% (95% CI, 52.9–84.7). The corresponding 2 × 2 table is shown in Table 3.

Table 3. Severe prostate enlargement (≥50 mL): per patient contingency table for the index test vs reference. AI positive was defined as AI-estimated prostate volume ≥50 mL; reference positive was defined as manual planimetry (ellipsoid formula) ≥50 mL. Sensitivity, specificity, positive predictive value, and negative predictive value were reported with 95% CIs calculated using the Wilson method. AI, artificial intelligence; CI, confidence interval.
Reference
≥50 mL<50 mL
AI≥50 mL56
<50 mL215

The agreement with the reference was ICC(A,1) = 0.789, indicating moderate absolute agreement between a single AI measurement and the expert reference. The correlation coefficient was Pearson’s r = 0.821 (95% CI: 0.647–0.914) (Figure 4). Bland–Altman analysis showed a mean bias of +5.1 mL (AI minus reference) and relatively wide 95% limits of agreement from −25.0 to +35.1 mL (Figure 5), indicating substantial individual-level variability despite acceptable group-level correlation. Representative examples of AI underestimation and overestimation of prostate volume are shown in Figure 6.

Figure 4. AI-estimated vs reference prostate volume (per patient). Scatterplot of AI-estimated versus reference prostate volume; the diagonal line represents y = x (perfect agreement). Pearson’s r = 0.821 (95% CI, 0.65–0.91). CI, confidence interval; AI, artificial intelligence.
Figure 5. Bland–Altman plot for prostate volume (per patient). Differences (AI − reference, mL) are plotted against the mean volume (mL). The middle horizontal line indicates the mean bias (+5.1 mL), and the upper and lower horizontal lines indicate the 95% limits of agreement (−25.0 to +35.1 mL; bias ± 1.96 SD). ICC(A,1) = 0.789 (two-way mixed-effects, absolute-agreement, single-measure). ICC, intraclass correlation coefficient; SD, standard deviation; AI, artificial intelligence.
Figure 6. Examples of AI under- and overestimation of prostate volume on handheld ultrasound. Four nurse-acquired suprapubic ultrasound images illustrate discrepancies between AI-estimated and reference prostate volume. (A–B) Underestimation by AI relative to the expert reference; (C–D) overestimation. (A, C) Transverse views with red lines indicating left–right width; (B, D) sagittal views with blue lines indicating anteroposterior depth and green lines indicating cranio-caudal length. Thick colored lines indicate AI-estimated diameters, whereas thin colored lines indicate expert reference measurements. AI, artificial intelligence.

Discussion

In this pilot diagnostic accuracy study, we developed an AI system tailored to handheld POCUS and evaluated it using nurse-acquired scans. The model showed high sensitivity with acceptable specificity for hydronephrosis detection. For prostate volume, the model showed moderate agreement with the expert reference and may have potential as a triage-support tool, but individual-level errors were substantial, as reflected by the wide Bland–Altman limits of agreement.

Our focus on nurse acquisition aligns with prior work showing that nurses can be trained to obtain urological ultrasound and integrate the results into care pathways in home and ward settings [10]. Integrative reviews indicate that registered nurses and nurse practitioners can reliably perform POCUS with appropriate training, supporting an expanded imaging role at the bedside [20]. Quality improvement and implementation reports of nurse-performed bladder ultrasound demonstrate its feasibility and perceived efficiency gains in routine care [21]. Nurse-led LUTS clinics have reduced specialist workloads without compromising safety [22]. Beyond feasibility, bladder scanners have been associated with fewer unnecessary catheterizations and reduced catheter-associated urinary tract infection (UTI), highlighting the tangible patient benefits of nurse-integrated ultrasound workflows [23].

The novelty of our contribution lies in the prospective evaluation of AI on handheld (pocket) POCUS using nurse-acquired clips after brief instruction, thereby testing the system under the image-quality and operator-variability constraints inherent to point-of-care nursing workflows. By prespecifying operating points aligned to triage/decision support, our approach emphasizes practical deployment. Overall, this extends prior AI efforts—largely console-based or expert-acquired—and provides nurse-validated evidence.

This clinical context supports our choice of the operating point for hydronephrosis. In non-expert hands, POCUS typically yields moderate sensitivity and specificity compared to computed tomography (CT) (often approximately 70–80% each) [24–26]. A randomized trial showed that ultrasound-first imaging can reduce radiation exposure for suspected nephrolithiasis without worsening safety outcomes. However, the performance varies with operator experience and case severity [26]. Against this backdrop, our high sensitivity and correspondingly high NPV at a fixed threshold support a rule-out role in nurse-led workflows. When the AI output is negative, clinically important dilatation is unlikely, permitting de-escalation or alternative diagnoses while reserving confirmatory imaging for positive or equivocal cases [24–26].

For the prostate, conventional ellipsoid formulas are convenient yet can systematically misestimate true gland size. Contemporary series against surgical or advanced imaging references report underestimation with the standard 0.52 coefficient and improved accuracy with revised coefficients or “bullet”-shape assumptions [27]. Recent comparative data indicate strong correlations between imaging-based volume and specimen volume, with magnetic resonance imaging (MRI) generally outperforming transabdominal ultrasonography (TAUS) for large glands. This highlights the challenges of ultrasound-based volumetry [28]. In this context, modern deep learning approaches to ultrasound autosegmentation have shown promising reproducibility and accuracy for prostate boundary delineation and derived volumes, offering a pathway to reduce operator dependence [29]. Consistent with these trends, our nurse-validated system showed moderate agreement with expert planimetry and modest screening performance at the prespecified ≥50 mL threshold. The wide limits of agreement and the observed under- and overestimation cases indicate that the system should be considered a triage-support aid rather than a definitive volumetric measurement method.

This study had several limitations. First, this was a single-center evaluation with a modest sample size, wide confidence intervals, and limited precision. Second, only one kidney per patient was included in the primary hydronephrosis analysis. Although this approach avoided within-patient correlation in a small pilot cohort, selecting the kidney with the clearest acoustic window may have introduced selection bias. This may overestimate performance compared with bilateral screening. Future studies should evaluate both kidneys and account for clustering within patients. Third, the model was designed for binary hydronephrosis detection and did not explicitly distinguish mild pelviectasis/renal pelvic fullness from clinically significant hydronephrosis or grade hydronephrosis severity. Future studies should evaluate severity categories such as mild, moderate, and severe hydronephrosis. Fourth, the reference for prostate volume was single-reader planimetry on the same acquisition rather than an independent modality, such as TAUS/MRI/specimen, or a multi-reader consensus. This limits the assessment of inter-reader variability and absolute volumetric accuracy. This underscores how reference choice affects apparent bias for ellipsoid-based estimates [28,29]. Fifth, five nurse-acquired prostate studies were non-diagnostic and excluded from accuracy analyses. Although this was prespecified because a reliable reference measurement could not be obtained, excluding non-diagnostic cases may overestimate operational performance among all attempted scans. The non-diagnostic rate should therefore be considered an implementation outcome. Sixth, bladder volume was not standardized or systematically recorded. Small bladder volumes may have limited the suprapubic acoustic window and contributed to prostate volume errors. Future protocols should standardize bladder filling or incorporate minimum bladder-volume and image-quality criteria. Seventh, participants had a relatively low BMI (prostate cohort median, 22.4 kg/m²; IQR, 21.0–24.0), and generalizability to patients with obesity or more difficult acoustic windows remains uncertain. Eighth, operator factors and image quality affect usability. Although scans were acquired by ward nurses after brief instructions, operator-level metadata—such as years of experience and prior POCUS use—were not collected, and we did not stratify the diagnostic performance of individual nurses. Therefore, potential interoperator differences and learning effects could not be assessed. Finally, the development of the model mixed conventional console and handheld data from one institution. Prospective validation on nurse scans improved the ecological validity, but generalizability to other devices, settings, and nurse cohorts remains to be confirmed.

Clinically, the hydronephrosis results support bedside rule-out aids operated by nurses, complementing ultrasound-first pathways that aim to reduce radiation exposure without compromising safety [24–26]. For prostate enlargement, the results should be interpreted as feasibility evidence for triage support rather than as definitive volumetry. Shape-prior or contour-completion approaches may help estimate prostate boundaries when anatomical delineation is incomplete and may reduce non-diagnostic rates in future models. Future work should pursue multicenter, multi-device external validation in diverse nurse cohorts, bilateral kidney evaluation with clustered analyses, multi-reader references, and, where feasible, independent comparators (TAUS/MRI/specimen), evaluation of impact outcomes (time-to-decision, catheterization rates, avoidable transfers), standardized bladder filling for prostate scans, real-time acquisition coaching, and case-mix-specific threshold calibration. The broader literature suggests that nurse–AI collaboration can improve consistency and timeliness of urinary tract assessments at the point of care [16,20,30–31].

Conclusions

Our nurse-validated evidence showed that AI models can be used to detect hydronephrosis and estimate prostate volume on handheld POCUS at prespecified operating points for triage/decision-support. This advances AI ultrasound from expert, console-based contexts to nurse-operated bedside care. However, the prostate volume results should be interpreted as feasibility evidence requiring further validation because non-diagnostic scans, wide limits of agreement, and individual-level errors remain important limitations.

Acknowledgments

We sincerely thank the patients who participated in this study. We are deeply grateful to the faculty and staff of the Department of Urology, the urology ward nurses, and the Clinical Laboratory at The University of Tokyo Hospital for their invaluable support. We also thank the many colleagues and related personnel across the hospital and the Graduate School of Medicine, including former affiliates and collaborators, who contributed to patient coordination and scheduling, image acquisition, data handling, and regulatory procedures; this study would not have been possible without their collaboration.

Ethics Statement

This study was approved by the Ethics Committee of the Graduate School of Medicine, The University of Tokyo (approval nos. 2019051NI-(3) and 2020301NI-(1)). All imaging data were de-identified at acquisition and used exclusively for research purposes. The study adhered to the Declaration of Helsinki and relevant institutional guidelines. For the prospective evaluation cohort, all participants received written information and provided written informed consent prior to imaging. For the retrospective development dataset, an IRB-approved opt-out procedure was implemented. Additionally, a detailed overview of the study was posted on the institutional website, and patients were allowed to decline participation at any point in the study.

Disclosure Statement

Masaru Matsumoto, Mikihiko Karube, Nao Tamai, and Yuka Miura were affiliated with a social collaboration department that received funding from Fujifilm Corporation through the end of fiscal year 2022 (FY2022). The other authors have disclosed no competing interests.

Funding

This work was supported in part by the joint research fund from the Department of Imaging Nursing Science, Graduate School of Medicine, The University of Tokyo. This study was also partly supported by JSPS KAKENHI Grant Number 21H03262 (Grant-in-Aid for Scientific Research (B) for M. Matsumoto). The funders had no role in the study design, data collection, analysis, decision to publish, or manuscript preparation.

Author Contributions

MaM: conceptualization; methodology; validation; formal analysis; data curation; writing – original draft; writing – review & editing; visualization; project administration. MoS: conceptualization; methodology; resources; writing – review & editing; visualization. MK: methodology; software; validation; formal analysis; data curation; writing – review & editing; visualization. HG: methodology; investigation; resources; writing – review & editing. MaS: resources; writing – review & editing. TK: investigation; writing – review & editing. MiM: investigation; writing – review & editing. SW: investigation; writing – review & editing. AK: validation; formal analysis; writing – review & editing. YM: conceptualization; writing – review & editing. NT: conceptualization; supervision; project administration; funding acquisition; writing – review & editing. HS: conceptualization; resources; supervision; funding acquisition; writing – review & editing. All authors read and approved the final manuscript.

Artificial Intelligence Statement

Generative AI tools were used for limited language editing and proofreading of sections of the abstract and main text. All authors reviewed and edited the output, take full responsibility for the content, and confirm that no generative AI tool was used to generate scientific results, analyze data, or draw conclusions.

Data Availability Statement

Availability of data and material: The clinical imaging datasets generated and/or analyzed during the current study are not publicly available and cannot be shared outside the institution due to ethics approvals and institutional policies governing patient data. Summary results are reported in the manuscript; no individual-level imaging data will be provided. The study code and trained model are not publicly available at this time because of institutional and ongoing development considerations; selected analysis details may be made available from the corresponding author upon reasonable request, subject to institutional approval.

References

1. Osuga Y, Okamura K, Ando F, Shimokata H. Prevalence of lower urinary tract symptoms in middle-aged and elderly Japanese. Geriatr Gerontol Int. 2013;13(4):1010-1017. doi:10.1111/ggi.12048.

2. Cabinet Office, Government of Japan. Annual Report on the Ageing Society: Summary 2021. 2021. Available from: https://www8.cao.go.jp/kourei/english/annualreport/index-wh.html. Accessed September 1, 2026.

3. Grzybowska ME, Wydra D. 24/7 usage of continence pads and quality of life impairment in women with urinary incontinence. Int J Clin Pract. 2019;73(8):e13267. doi:10.1111/ijcp.13267.

4. Kelleher C. Quality of life and urinary incontinence. Baillieres Best Pract Res Clin Obstet Gynaecol. 2000;14(2):363-379. doi:10.1053/beog.1999.0079.

5. Abdelmoteleb H, Jefferies ER, Drake MJ. Assessment and management of male lower urinary tract symptoms (LUTS). Int J Surg. 2016;25:164-171. doi:10.1016/j.ijsu.2015.11.043.

6. Ross SS, Forster CS, Borawski K. Urinary tract infection and neuropathic bladder. Urol Clin North Am. 2024;51(4):551-559. doi:10.1016/j.ucl.2024.06.009.

7. Zambon JP, Koslov DS, Mihai B, Badlani GH. Bladder and ureteral dysfunction leading to hydronephrosis and hydroureteronephrosis in adults. Urology. 2018;117:1-8. doi:10.1016/j.urology.2017.11.049.

8. Homma Y, Gotoh M, Kawauchi A, Kojima Y, Masumori N, Nagai A, Saitoh T, Sakai H, Takahashi S, Ukimura O, Yamanishi T, Yokoyama O, Yoshida M, Maeda K. Clinical guidelines for male lower urinary tract symptoms and benign prostatic hyperplasia. Int J Urol. 2017;24(10):716-729. doi:10.1111/iju.13401.

9. Takahashi S, Takei M, Asakura H, Gotoh M, Ishizuka O, Kato K, Koyama M, Takeyama M, Tomoe H, Yamanishi T, Yokoyama O, Yoshida M, Yoshimura Y, Yoshizawa T. Clinical Guidelines for Female Lower Urinary Tract Symptoms (second edition). Int J Urol. 2021;28(5):474-492. doi:10.1111/iju.14492.

10. Matsumoto M, Tamai N, Miura Y, Okawa Y, Yoshida M, Igawa Y, Nakagami G, Sanada H. Evaluation of a point-of-care ultrasound educational program for nurse educators. J Contin Educ Nurs. 2021;52(8):375-381. doi:10.3928/00220124-20210714-07.

11. Matsumoto M, Tsutaoka T, Yabunaka K, Handa M, Yoshida M, Nakagami G, Sanada H. Development and evaluation of automated ultrasonographic detection of bladder diameter for estimation of bladder urine volume. PLoS One. 2019;14(9):e0219916. doi:10.1371/journal.pone.0219916.

12. Matsumoto M, Tsutaoka T, Nakagami G, Tanaka S, Yoshida M, Miura Y, Sugama J, Okada S, Ohta H, Sanada H. Deep learning-based classification of rectal fecal retention and analysis of fecal properties using ultrasound images in older adult patients. Jpn J Nurs Sci. 2020;17(4):e12340. doi:10.1111/jjns.12340.

13. Natali T, Kurucz LM, Fusaglia M, Mertens LS, Ruers TJM, van Leeuwen PJ, Dashtbozorg B. Automatic prostate volume estimation in transabdominal ultrasound images. Eur J Radiol. 2025;191:112274. doi:10.1016/j.ejrad.2025.112274.

14. Albayrak NB, Akgul YS. Estimation of the prostate volume from abdominal ultrasound images by image-patch voting. Appl Sci. 2022;12(3):1390. doi:10.3390/app12031390.

15. Lien WC, Chang YC, Chou HH, Lin LC, Liu YP, Liu L, Chan YT, Kuan FS. Detecting hydronephrosis through ultrasound images using state-of-the-art deep learning models. Ultrasound Med Biol. 2023;49(3):723-733. doi:10.1016/j.ultrasmedbio.2022.10.001.

16. Alexa R, Kranz J, Kramann R, Kuppe C, Sanyal R, Hayat S, Casas Murillo LF, Hajili T, Hoffmann M, Saar M. Harnessing artificial intelligence for enhanced renal analysis: automated detection of hydronephrosis and precise kidney segmentation. Eur Urol Open Sci. 2024;62:19-25. doi:10.1016/j.euros.2024.01.017.

17. van den Kroonenberg DL, Delberghe FT, Jager A, Postema AW, Beerlage HP, Zwart W, Mischi M, Oddens JR. Development and validation of an algorithm for segmentation of the prostate and its zones from three-dimensional transrectal multiparametric ultrasound images. Eur Urol Open Sci. 2025;75:48-54. doi:10.1016/j.euros.2025.03.005.

18. Bossuyt PM, Reitsma JB, Bruns DE, Gatsonis CA, Glasziou PP, Irwig L, Lijmer JG, Moher D, Rennie D, de Vet HCW, Kressel HY, Rifai N, Golub RM, Altman DG, Hooft L, Korevaar DA, Cohen JF; STARD Group. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527. doi:10.1136/bmj.h5527.

19. Xie S, Girshick R, Dollar P, Tu Z, He K. Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017. p. 1492-1500. doi:10.1109/CVPR.2017.634.

20. Totenhofer R, Luck L, Wilkes L. Point of care ultrasound use by registered nurses and nurse practitioners in clinical practice: an integrative review. Collegian. 2021;28(4):456-463. doi:10.1016/j.colegn.2020.10.002.

21. Liptrott SJ, Saliquni A, Giau L, Pecora R, Cimino C, Lo Piccolo A, Palermo A, Vidovic B, Cheda E, Imhof S, Tolotti A, Guastamacchia I, Cavatorti S. Nurse performed bladder ultrasound: a clinical quality improvement initiative. Contemp Nurse. 2025;61(3):270-286. doi:10.1080/10376178.2024.2406350.

22. Keane KG, Inder MS, McIntyre C, Omer S, McEvoy E, Smyth LG, Casey RG, Thomas AZ, Manecksha RP, Flynn RJ. Implementation of a nurse-led lower urinary tract symptoms (LUTS) clinic reduces general urology clinic workload in a Model 4 Hospital: a pilot study in Tallaght University Hospital. Ir J Med Sci. 2021;190(3):1123-1128. doi:10.1007/s11845-020-02428-8.

23. Palese A, Buchini S, Deroma L, Barbone F. The effectiveness of the ultrasound bladder scanner in reducing urinary tract infections: a meta-analysis. J Clin Nurs. 2010;19(21-22):2970-2979. doi:10.1111/j.1365-2702.2010.03281.x.

24. Herbst MK, Rosenberg G, Daniels B, Gross CP, Singh D, Molinaro AM, Luty S, Moore CL. Effect of provider experience on clinician-performed ultrasonography for hydronephrosis in patients with suspected renal colic. Ann Emerg Med. 2014;64(3):269-276. doi:10.1016/j.annemergmed.2014.01.012.

25. Sibley S, Roth N, Scott C, Rang L, White H, Sivilotti MLA, Bruder E. Point-of-care ultrasound for the detection of hydronephrosis in emergency department patients with suspected renal colic. Ultrasound J. 2020;12(1):31. doi:10.1186/s13089-020-00178-3.

26. Smith-Bindman R, Aubin C, Bailitz J, Bengiamin RN, Camargo CA Jr, Corbo J, Dean AJ, Goldstein RB, Griffey RT, Jay GD, Kang TL, Kriesel DR, Ma OJ, Mallin M, Manson W, Melnikow J, Miglioretti DL, Miller SK, Mills LD, Miner JR, Moghadassi M, Noble VE, Press GM, Stoller ML, Valencia VE, Wang J, Wang RC, Cummings SR. Ultrasonography versus computed tomography for suspected nephrolithiasis. N Engl J Med. 2014;371(12):1100-1110. doi:10.1056/NEJMoa1404446.

27. Aprikian S, Luz M, Brimo F, Scarlata E, Hamel L, Cury FL, Tanguay S, Aprikian AG, Kassouf W, Chevalier S. Improving ultrasound-based prostate volume estimation. BMC Urol. 2019;19(1):68. doi:10.1186/s12894-019-0492-2.

28. Guo S, Zhang J, Jiao J, Li Z, Wu P, Jing Y, Qin W, Wang F, Ma S. Comparison of prostate volume measured by transabdominal ultrasound and MRI with the radical prostatectomy specimen volume: a retrospective observational study. BMC Urol. 2023;23(1):62. doi:10.1186/s12894-023-01234-5.

29. Kurucz LM, Natali T, Fusaglia M, Dashtbozorg B. Advances in deep learning methods for prostate segmentation and volume estimation in ultrasound imaging. Appl Sci. 2024;14(15):6550. doi:10.3390/app14156550.

30. Gao Y, Zeng S, Xu X, Li H, Yao S, Song K, Li X, Chen L, Tang J, Xing H, Yu Z, Zhang Q, Zeng S, Yi C, Xie H, Xiong X, Cai G, Wang Z, Wu Y, Chi J, Jiao X, Qin Y, Mao X, Chen Y, Jin X, Mo Q, Chen P, Huang Y, Shi Y, Wang J, Zhou Y, Ding S, Zhu S, Liu X, Dong X, Cheng L, Zhu L, Cheng H, Cha L, Hao Y, Jin C, Zhang L, Zhou P, Sun M, Xu Q, Chen K, Gao Z, Zhang X, Ma Y, Liu Y, Xiao L, Xu L, Peng L, Hao Z, Yang M, Wang Y, Ou H, Jia Y, Tian L, Zhang W, Jin P, Tian X, Huang L, Wang Z, Liu J, Fang T, Yan D, Cao H, Ma J, Li X, Zheng X, Lou H, Song C, Li R, Wang S, Li W, Zheng X, Chen J, Li G, Chen R, Xu C, Yu R, Wang J, Xu S, Kong B, Xie X, Ma D, Gao Q. Deep learning-enabled pelvic ultrasound images for accurate diagnosis of ovarian cancer in China: a retrospective, multicentre, diagnostic study. Lancet Digit Health. 2022;4(3):e179-e187. doi:10.1016/S2589-7500(21)00278-8.

31. Xu T, Zhang XY, Yang N, Jiang F, Chen GQ, Pan XF, Peng YX, Cui XW. A narrative review on the application of artificial intelligence in renal ultrasound. Front Oncol. 2024;13:1252630. doi:10.3389/fonc.2023.1252630.

Posted in .

Leave a Reply

Your email address will not be published. Required fields are marked *