Deep Learning‐Assisted Classification of Urinary Red Blood Cell Morphology for Glomerular Hematuria Screening: A Pilot Study

ABSTRACT

Background

Distinguishing glomerular from non‐glomerular hematuria remains challenging because urinary dysmorphic red blood cells (RBCs) are morphologically heterogeneous and affected by preanalytical and physicochemical factors. This pilot study developed a deep learning‐assisted system for urinary RBC morphology classification and evaluated its feasibility for expert‐guided glomerular hematuria screening support.

Methods

We retrospectively analyzed 491 high‐resolution urine sediment images containing 15,779 annotated RBCs or RBC‐like objects from a regional teaching hospital. RBCs were labeled as isomorphic, dysmorphic, or unknown according to established morphological criteria. A YOLOv5l model was trained for RBC detection and classification. Model outputs were integrated with an operational dysmorphic RBC‐based scoring system and compared with manual expert assessment.

Results

The model achieved a precision of 0.84, recall of 0.69, and F1‐score of 0.76 for dysmorphic RBC classification, with a recall of 0.98 for isomorphic RBCs. In sample‐level scoring, concordance with expert assessment was 100% in the Negative category (39/39; 95% CI, 91.0%–100%), 81.8% in the Moderate category (9/11; 95% CI, 48.2%–97.7%), and 78.1% in the Major category (25/32; 95% CI, 60.0%–90.7%). No validation samples were available in the Few category. Mean computational inference time was 0.033 s per image.

Conclusions

This YOLOv5l‐based pilot study demonstrates the feasibility of rapid, morphology‐aware urinary RBC classification and preliminary concordance with expert microscopy. The system may support expert‐guided workflow research, but should not be interpreted as standalone diagnostic performance. Multicenter validation with independent clinical reference standards is required before clinical implementation.

This retrospective pilot study developed a YOLOv5l‐based deep learning system to detect and classify urinary red blood cells as isomorphic, dysmorphic, or unknown in urine sediment images. The model achieved a precision of 0.84, recall of 0.69, and F1‐score of 0.76 for dysmorphic RBCs, while sample‐level morphology scoring showed preliminary concordance with expert microscopic assessment. These findings support the feasibility of rapid, morphology‐aware urinary RBC classification as an expert‐guided laboratory workflow tool, although external multicenter validation and independent clinical reference standards are still required.

1. Introduction

Hematuria is a common clinical finding associated with a broad spectrum of urinary tract and kidney disorders, ranging from benign urological conditions to glomerulonephritis, IgA nephropathy, and other intrinsic renal diseases. Differentiating glomerular hematuria from non‐glomerular hematuria is clinically important because it helps guide subsequent nephrological or urological evaluation. As the classic observation that urinary red blood cell (RBC) morphology can help distinguish glomerular from non‐glomerular bleeding, microscopic examination of urinary sediment has remained an important laboratory approach for evaluating the origin of hematuria [1]. In glomerular hematuria, RBCs undergo structural deformation during passage through the damaged glomerular filtration barrier and renal tubules, whereas non‐glomerular bleeding is typically associated with more uniform, isomorphic RBCs [2]. Persistent glomerular hematuria has also been associated with an increased risk of progressive kidney disease, further emphasizing the need for accurate and reproducible laboratory assessment [3, 4].

Despite its clinical relevance, the interpretation of urinary RBC morphology is biologically complex. Dysmorphic RBCs are not a single uniform entity but comprise multiple morphological patterns, including acanthocytes, ghost cells, fragmented cells, and other irregular forms. Among these, acanthocytes, also referred to as G1 cells in some classification systems, are considered one of the most specific morphological markers of glomerular bleeding because of their characteristic ring‐shaped appearance and membrane protrusions [4, 5]. However, not all dysmorphic RBCs are equally specific for glomerular injury, and some abnormal forms may arise from non‐glomerular conditions or preanalytical influences [6, 7]. Therefore, any automated approach for urinary RBC classification must be grounded in clearly defined morphological criteria and expert‐based annotation rather than treating the task as a purely computational detection problem.

Manual urine sediment microscopy remains the reference method for identifying urinary RBC morphology, but it is time‐consuming, operator‐dependent, and subject to inter‐observer variability [7, 8]. The reproducibility of dysmorphic RBC assessment depends on several factors, including specimen freshness, centrifugation and sediment preparation, microscope settings, observer expertise, and the criteria used to define dysmorphic forms [7, 9]. In addition, urine osmolality, specific gravity, and pH may substantially influence RBC appearance. Hypotonic urine can cause RBC swelling, hemolysis, and ghost‐cell formation, whereas hypertonic urine may induce cell shrinkage or crenation, potentially mimicking dysmorphic features. These osmotic and preanalytical effects are important confounders for both manual microscopy and image‐based artificial intelligence systems because they directly alter the visual features used for classification.

Automated urine particle analyzers, including flow cytometry‐based systems such as the Sysmex UF series, have improved the throughput and standardization of routine urinalysis [10]. These systems can provide useful information on hematuria and related urinary particle parameters; however, their assessment of glomerular hematuria is generally based on indirect physical properties, such as cell size, forward scatter, or urinary RBC distribution, rather than direct recognition of membrane protrusions or acanthocytic morphology [11, 12]. Recent multicenter evidence suggests that urinary RBC distribution may offer comparable diagnostic performance to dysmorphic RBC assessment in some clinical settings, but it remains a surrogate parameter and does not fully replace morphology‐based microscopic evaluation [12]. Therefore, a morphology‐aware image analysis method may provide complementary information by directly learning visual RBC features from urine sediment images.

Deep learning has shown increasing utility in medical image analysis because convolutional neural networks can learn complex morphological patterns directly from image data [13, 14]. Among object detection architectures, the You Only Look Once version 5 (YOLOv5) framework offers rapid inference and has been applied in several biomedical image recognition tasks, including cellular and microscopic object detection [15, 16]. However, the application of YOLO‐based deep learning to urinary RBC morphology remains underexplored. In particular, previous studies have rarely addressed whether automated models can classify urinary RBCs in a way that reflects the biological heterogeneity of dysmorphic RBCs and the practical requirements of clinical laboratory workflows.

In this pilot study, we developed and evaluated a YOLOv5l‐based deep learning model for urinary RBC morphology classification using high‐resolution urine sediment images. RBCs were classified into isomorphic, dysmorphic, and unknown categories based on established morphological criteria, with attention to dysmorphic features relevant to glomerular hematuria. By integrating model outputs with a dysmorphic RBC‐based scoring system, we aimed to assess the feasibility of a morphology‐aware deep learning approach for glomerular hematuria screening and laboratory workflow support rather than to replace expert microscopic interpretation.

2. Materials and Methods

2.1. Study Design and Image Dataset

This retrospective image‐based pilot study was conducted at Pingtung Christian Hospital, a regional teaching hospital in Taiwan. Urine sediment images were collected from the routine clinical laboratory workflow between January 1, 2021, and September 30, 2022. The dataset comprised urine samples from patients with macroscopic or microscopic hematuria identified during routine urinalysis.

The study protocol was reviewed and approved by the Institutional Review Board of Pingtung Christian Hospital (IRB No. IRB731A). Because this study used retrospective, de‐identified images and laboratory data without additional patient intervention, the requirement for informed consent was waived by the Institutional Review Board.

A total of 491 high‐resolution urine sediment images were included. These images contained 15,779 annotated urinary red blood cells (RBCs) or RBC‐like objects, which were used for model training and validation. To reduce the risk of data leakage, the dataset was partitioned at the urine‐sample level. All images derived from the same urine sample were assigned exclusively to either the training set or the validation set. This resulted in an approximately 8:2 image‐level split, with 409 images in the training set and 82 images in the validation set. The validation set comprised 82 images derived from 82 distinct urine samples, with one image corresponding to each sample. Thus, the image‐level and urine‐sample‐level units were equivalent within the validation set. Because the retrospective dataset was de‐identified and organized primarily by urine specimen, longitudinal patient‐level separation could not be fully evaluated and was therefore considered a limitation of the study.

2.2. Urine Sample Preparation and Image Acquisition

Urine samples were processed according to a modified urine sediment preparation protocol designed to enhance the recovery of formed elements, including dysmorphic RBCs and renal tubular epithelial cells [17]. Briefly, 10 mL of urine was centrifuged at 500 g for 5 min. After centrifugation, the supernatant was carefully discarded, leaving approximately 0.6 mL of sediment for resuspension.

The resuspended sediment was mixed gently, and a drop of the sediment suspension was placed on a glass slide for microscopic examination. High‐resolution images were captured using a Nikon Eclipse Ci‐L microscope equipped with a Nikon‐view 8000AIO‐Lite imaging system. Image acquisition was performed under standardized laboratory conditions to minimize variation in illumination, magnification, and focus.

Because urinary RBC morphology may be affected by preanalytical and physicochemical factors, including urine osmolality, specific gravity, and pH, particular attention was given during image review to osmotic artifact‐related changes [7, 9]. These changes included ghost cells, swollen RBCs, crenated cells, and severely distorted or overlapping cells. Objects that could not be reliably classified as isomorphic or dysmorphic RBCs were assigned to the “unknown” category.

2.3. Morphological Annotation Criteria

Manual image annotation was performed using LabelMe, a web‐based image annotation tool [18]. Individual RBCs or RBC‐like objects were annotated with bounding boxes and classified into one of three categories: isomorphic RBCs, dysmorphic RBCs, or unknown objects.

The annotation criteria were based on established urinary RBC morphological classifications and sediment microscopy guidelines [5, 6, 7, 9]. Isomorphic RBCs were defined as RBCs with relatively uniform size and shape, smooth contours, and preserved round or biconcave morphology, consistent with non‐glomerular bleeding. Dysmorphic RBCs were defined as RBCs showing irregular contours, marked variation in size or shape, membrane abnormalities, or protrusion‐like structures. When identifiable, acanthocytic or G1‐like features, including ring‐shaped cells with vesicle‐like membrane protrusions, were considered representative dysmorphic features suggestive of glomerular origin [5, 6].

The “unknown” category was used for cells or objects with ambiguous morphology, severe overlap, poor focus, partial cell borders, marked hemolysis, ghost‐like appearance, or background artifacts that could not be confidently assigned to either the isomorphic or dysmorphic category. This category was included to avoid forcing biologically or technically ambiguous objects into clinically meaningful RBC classes and to improve the reproducibility of annotation. The final annotated dataset included 2312 dysmorphic RBCs, 12,597 isomorphic RBCs, and 870 unknown objects.

2.4. Annotation Reproducibility Assessment

To assess the reproducibility of expert urinary RBC morphology assessment, two complementary interobserver agreement analyses were performed. First, sample‐level reproducibility was evaluated using the 82 validation images, each of which represented a distinct urine sample. A second experienced medical technologist independently reviewed the image corresponding to each sample without access to the AI predictions and assigned the sample to one of the operational dysmorphic RBC‐based morphology scoring categories: Negative (< 15% dysmorphic RBCs), Few (15%–29%), Moderate (30%–59%), or Major (≥ 60%). Agreement between the original expert assessment and the second independent review was evaluated using Cohen's kappa. Because these scoring categories are ordinal, linearly weighted Cohen's kappa was also calculated.

Second, object‐level annotation reproducibility was assessed using a stratified random subset of 500 RBC or RBC‐like objects selected from the annotated dataset. To ensure adequate representation of each morphology class, the subset included 200 isomorphic RBCs, 200 dysmorphic RBCs, and 100 unknown objects. These objects were independently re‐annotated by the second experienced medical technologist as isomorphic RBC, dysmorphic RBC, or unknown. Object‐level interobserver agreement was evaluated using Cohen's kappa. These reproducibility analyses were performed to assess the reliability of expert morphology assessment and were not used for model training.

2.5. Deep Learning Model Architecture and Training

A YOLOv5l model was used for urinary RBC detection and morphology classification. YOLOv5l was selected because it provides a practical balance between detection performance and computational efficiency and has a stable, widely used implementation suitable for pilot biomedical object‐detection studies. Although newer object‐detection architectures have become available, YOLOv5l was considered appropriate for this feasibility study because of its reproducibility, implementation stability, and ability to perform rapid inference on microscopy images.

The model was implemented using the Ultralytics YOLOv5 v7.0 framework in a Python 3.10.12 and PyTorch environment. Training was performed on a Tesla T4 GPU with 15 GB of available GPU memory in a computing environment with 12.7 GB of system RAM. Images and annotations generated using LabelMe were converted into the YOLO‐compatible object‐detection format before training. The model was trained for 100 epochs using a batch size of 16, stochastic gradient descent optimization, and an initial learning rate of 0.01. The training IoU threshold was set at 0.20. For inference, the confidence threshold was set at 0.25, and the non‐maximum suppression IoU threshold was set at 0.45.

Before model training, urine sediment images were resized to 640 × 640 pixels using the standard YOLOv5 v7.0 preprocessing pipeline. During model training, the data‐augmentation strategy included mosaic augmentation and image scaling. The three annotation classes, isomorphic RBC, dysmorphic RBC, and unknown, were retained during training to allow the model to distinguish clear RBC morphologies from ambiguous or artifact‐like objects.

The annotated dataset showed class imbalance, with isomorphic RBCs being substantially more frequent than dysmorphic RBCs and unknown objects. The final dataset included 12,597 isomorphic RBCs, 2312 dysmorphic RBCs, and 870 unknown objects. This distribution reflected the relative availability of RBC morphologies in the retrospective urine sediment image dataset. No dedicated class‐balancing strategy, such as oversampling, class‐weighted loss, or focal loss, was applied beyond the specified YOLOv5 training pipeline. Therefore, the potential influence of class imbalance on model performance, particularly dysmorphic RBC recall, was considered during interpretation of the results.

2.6. Dysmorphic RBC‐Based Morphology Scoring

To evaluate sample‐level applicability, model‐predicted RBC classifications were integrated into an operational dysmorphic RBC‐based morphology scoring system. For each sample, the proportion of dysmorphic RBCs was calculated as follows:

$$\begin{array}{c}\text{Dysmorphic}\mathrm{RBC}\text{percentage}=\hfill \\ \text{dysmorphic}\mathrm{RBC}\text{count}/\text{total classifiable}\mathrm{RBC}\text{count}\times 100\%\hfill \end{array}$$

The total classifiable RBC count included RBCs assigned to either the isomorphic or dysmorphic category. Objects assigned to the unknown category were retained for model training and error analysis but were not used as confidently classifiable RBCs for sample‐level scoring.

The scoring thresholds were derived from a previously used dysmorphic RBC scoring template in which dysmorphic RBC counts are interpreted in relation to the total RBC count per microscopic field. To adapt this semi‐quantitative scoring template to image‐based AI outputs, the original count‐based scoring categories were converted into percentage ranges. Based on the lower‐bound percentage values of each category, the operational thresholds were defined as follows: Negative, < 15% dysmorphic RBCs; Few, 15%–29%; Moderate, 30%–59%; and Major, ≥ 60%. The operational conversion of these dysmorphic RBC scoring categories into percentage‐based thresholds is summarized in Table S1.

These categories were used as operational morphology‐based groups for feasibility assessment and comparison with manual expert assessment. They were not intended to serve as externally validated diagnostic thresholds for glomerular disease. Therefore, sample‐level scoring results were interpreted as concordance with expert microscopic assessment rather than as biopsy‐confirmed diagnostic performance.

2.7. Model Performance Evaluation

Model performance was evaluated on the validation dataset using object‐level detection and classification metrics. For each morphology class, true‐positive (TP), false‐positive (FP), and false‐negative (FN) counts were recorded. Precision, recall, and F1‐score were calculated from these counts. Precision was defined as the proportion of correctly predicted objects among all objects predicted as a given class, and recall was defined as the proportion of correctly predicted objects among all ground‐truth objects of that class. The F1‐score was calculated as the harmonic mean of precision and recall.

Because true‐negative counts are not routinely defined in the same way as TP, FP, and FN in object‐detection outputs, accuracy was not used as the primary object‐level performance metric in the revised analysis. In addition to object‐level performance, sample‐level concordance was evaluated by comparing AI‐based dysmorphic RBC scoring categories with manual expert‐assigned categories.

Representative misclassification patterns were also reviewed qualitatively to identify morphology‐related error sources. Particular attention was given to false‐positive isomorphic RBC predictions, dysmorphic RBCs misclassified as isomorphic RBCs or unknown objects, and background or artifact‐like structures misclassified as RBCs. Error patterns were interpreted in relation to common urine sediment image challenges, including overlapping cells, poor focus, ghost‐like cells, air bubbles, partial cell borders, and background debris.

2.8. Efficiency Analysis

To assess technical feasibility for rapid image‐level interpretation, the computational inference time of the YOLOv5l model was compared with manual microscopic interpretation time. Manual interpretation was performed by 10 medical technologists with varying levels of professional experience, and the time required for manual interpretation was recorded under routine laboratory conditions.

AI processing time was defined as computational inference time per image. This measurement did not include image acquisition, image loading, preprocessing, expert review of AI‐generated outputs, or quality‐control procedures. Therefore, the timing analysis was intended to evaluate computational inference speed rather than a complete end‐to‐end clinical workflow comparison.

2.9. Statistical Analysis

Descriptive statistics were used to summarize the dataset composition, including the number of images, annotated RBCs, and class distribution. Object‐level model performance was evaluated using TP, FP, and FN counts for each class. Precision, recall, and F1‐score were calculated from these counts. Exact binomial 95% confidence intervals were calculated for precision and recall. F1‐score was reported descriptively as a composite metric derived from precision and recall. Because true‐negative counts are not routinely defined in the same way as TP, FP, and FN in object‐detection outputs, accuracy was not used as the primary object‐level performance metric in the revised analysis.

Sample‐level concordance between AI‐based dysmorphic RBC scoring and manual expert assessment was expressed as the number and percentage of correctly classified samples within each operational morphology category. Exact binomial 95% confidence intervals were calculated for sample‐level concordance rates.

Interobserver agreement for expert morphology assessment was evaluated using Cohen's kappa. Because the Negative, Few, Moderate, and Major sample‐level scoring categories are ordinal, linearly weighted Cohen's kappa was additionally calculated. Ninety‐five percent confidence intervals for the unweighted and linearly weighted kappa statistics were derived using asymptotic standard errors. Interobserver agreement analyses were performed using IBM SPSS Statistics, version 26.0 (IBM Corp., Armonk, NY, USA).

Processing time was summarized as mean time per image or microscopic field. AI processing time was defined as computational inference time per image and did not include image acquisition, image loading, preprocessing, review of AI‐generated outputs, or quality‐control procedures. Because this was a retrospective pilot study focused on feasibility and morphology classification, no formal sample size calculation was performed. Diagnostic accuracy against biopsy‐confirmed glomerular disease or other independent clinical reference standards was not assessed. ROC/AUC analysis was not performed because complete object‐level prediction confidence scores required for threshold‐sweeping analysis were not retained in the archived retrospective object‐detection outputs.

3. Results

3.1. Dataset Characteristics and Morphological Annotation

A total of 491 high‐resolution urine sediment images were included in this retrospective pilot study. To reduce potential data leakage, the dataset was partitioned at the urine‐sample level, with all images derived from the same urine sample assigned exclusively to either the training set or the validation set. This resulted in an approximately 8:2 image‐level split, with 409 images in the training set and 82 images in the validation set. The 82 validation images represented 82 distinct urine samples, with one image corresponding to each sample. Overall, 15,779 urinary red blood cells (RBCs) or RBC‐like objects were manually annotated.

The annotated dataset included 12,597 isomorphic RBCs, 2312 dysmorphic RBCs, and 870 unknown objects. Isomorphic RBCs accounted for the largest proportion of annotated objects, whereas dysmorphic RBCs represented a smaller but clinically important subset. The unknown category included objects with ambiguous morphology, overlapping cells, poor focus, ghost‐like appearance, or background artifacts that could not be confidently assigned to either the isomorphic or dysmorphic RBC category. This three‐class annotation strategy was used to preserve morphologically uncertain objects rather than forcing them into clinically meaningful RBC categories.

3.2. Model Training Performance

The YOLOv5l model demonstrated stable convergence during training. As shown in Figure 1, the training and validation loss curves for bounding box regression, objectness, and classification decreased progressively over 100 epochs. In parallel, performance indicators, including precision, recall, mAP_0.5, and mAP_0.5:0.95, increased during training and reached a plateau.

FIGURE 1.

Model training performance over 100 epochs. Training and validation loss curves for bounding box regression, objectness, and classification decreased progressively, while precision, recall, mAP_0.5, and mAP_0.5:0.95 increased and reached a plateau, indicating stable model convergence.

These training patterns indicated that the model successfully learned visual features from the annotated urine sediment images without obvious divergence between the training and validation curves.

3.3. Object‐Level Classification Performance

The object‐level performance of the YOLOv5l model on the validation dataset is summarized in Table 1, and the corresponding confusion matrix is shown in Figure 2. For dysmorphic RBC classification, the model identified 381 true‐positive objects, with 75 false‐positive and 171 false‐negative predictions. This corresponded to a precision of 0.84 (95% CI, 0.80–0.87), recall of 0.69 (95% CI, 0.65–0.73), and F1‐score of 0.76.

TABLE 1.

Object‐level performance metrics of the YOLOv5l model for urinary RBC morphology classification.

Class	TP	FP	FN	Precision (95% CI)	Recall (95% CI)	F1‐score
Dysmorphic RBC	381	75	171	0.84 (0.80–0.87)	0.69 (0.65–0.73)	0.76
Isomorphic RBC	1148	1314	23	0.47 (0.45–0.49)	0.98 (0.97–0.99)	0.63
Unknown	65	58	44	0.53 (0.44–0.62)	0.60 (0.50–0.69)	0.56

Note: Exact binomial 95% confidence intervals were calculated for precision and recall. F1‐score is shown as a descriptive composite metric derived.

Abbreviations: FN, false negative; FP, false positive; RBC, red blood cell; TP, true positive.

FIGURE 2.

Confusion matrix of the YOLOv5l model on the validation dataset. The confusion matrix illustrates correct and incorrect object‐level predictions among isomorphic RBCs, dysmorphic RBCs, unknown objects, and background/artifact‐like regions. The large number of background or artifact‐like objects predicted as isomorphic RBCs contributed substantially to the low precision of the isomorphic RBC category.

For isomorphic RBCs, the model identified 1148 true‐positive objects and showed a high recall of 0.98 (95% CI, 0.97–0.99), indicating that most isomorphic RBCs in the validation dataset were successfully detected. However, the model generated 1314 false‐positive isomorphic RBC predictions, resulting in a precision of 0.47 (95% CI, 0.45–0.49) and an F1‐score of 0.63. As illustrated in Figure 2, background or artifact‐like objects contributed substantially to the false‐positive predictions and the lower precision of this category.

For the unknown category, the model identified 65 true‐positive objects, with 58 false‐positive and 44 false‐negative predictions. This corresponded to a precision of 0.53 (95% CI, 0.44–0.62), recall of 0.60 (95% CI, 0.50–0.69), and F1‐score of 0.56. These findings indicate that the model showed only moderate performance in identifying ambiguous or artifact‐like objects, reflecting the heterogeneous nature of the unknown category.

Because true‐negative counts are not routinely defined in the same way as TP, FP, and FN in object‐detection outputs, object‐level performance was primarily summarized using precision, recall, and F1‐score. These metrics were therefore used as the main quantitative performance indicators for object‐level urinary RBC morphology classification.

3.4. Morphology‐Related Classification Patterns and Error Sources

Representative morphology‐related classification patterns and error sources identified in the validation dataset are shown in Figure 3. The model correctly identified typical dysmorphic RBCs with clearly irregular contours and protrusion‐like morphology (Figure 3A). However, some dysmorphic RBCs were misclassified as isomorphic RBCs when the abnormal contour was relatively subtle or when the cell retained a partially preserved round appearance (Figure 3B). Other dysmorphic RBCs were classified as unknown when dysmorphic features were less distinct or morphologically ambiguous (Figure 3C).

FIGURE 3.

Representative morphology‐related classification patterns and error sources observed in the validation dataset. Green boxes indicate expert ground‐truth annotations, and red boxes indicate YOLOv5l model predictions. (A) Correctly identified dysmorphic RBC with a clearly irregular contour and protrusion‐like morphology. (B) Dysmorphic RBC misclassified as an isomorphic RBC. (C) Dysmorphic RBC classified as unknown because of ambiguous or less distinct dysmorphic features. (D) Background artifact misclassified as an isomorphic RBC. (E) Ghost‐like cell, air‐bubble‐like structure, or artifact‐like object classified as unknown. (F) Overlapping RBCs representing a challenging classification scenario, in which a dysmorphic RBC was misclassified as an isomorphic RBC. Panels D and E illustrate representative non‐definitive or artifact‐like objects identified during qualitative error review; therefore, only model predictions are shown. These examples demonstrate common sources of classification difficulty, including subtle dysmorphic morphology, artifact interference, ghost‐like structures, and overlapping cell boundaries.

False‐positive isomorphic RBC predictions were also observed in background artifact‐like structures that partially resembled RBC contours (Figure 3D). In addition, ghost‐like cells, air‐bubble‐like structures, and other artifact‐like objects contributed to uncertain classifications and were frequently assigned to the unknown category (Figure 3E). Overlapping RBCs represented another challenging scenario because superimposed cell boundaries could obscure dysmorphic features and lead to misclassification (Figure 3F).

Overall, these representative examples indicate that classification errors were associated not only with model limitations but also with the intrinsic morphological complexity of urine sediment images. Major error sources included subtle dysmorphic morphology, background artifacts, ghost‐like structures, and overlapping cell boundaries.

3.5. Interobserver Agreement for Urinary RBC Morphology Assessment

Interobserver agreement for expert urinary RBC morphology assessment is summarized in Table S2. For sample‐level dysmorphic RBC‐based morphology scoring, the original expert assessment and the second independent review showed 92.7% overall agreement across the 82 distinct validation samples, each represented by one urine sediment image. Cohen's kappa was 0.88 (95% CI, 0.78–0.98), indicating high agreement. Because the Negative, Few, Moderate, and Major categories are ordinal, linearly weighted Cohen's kappa was also calculated and showed almost perfect agreement (linearly weighted κ = 0.95; 95% CI, 0.90–0.99).

For object‐level annotation reproducibility, a stratified random subset of 500 RBC or RBC‐like objects was independently re‐annotated by a second experienced medical technologist. The subset included 200 isomorphic RBCs, 200 dysmorphic RBCs, and 100 unknown objects. The overall agreement for object‐level classification was 91.0%, with a Cohen's kappa of 0.82 (95% CI, 0.77–0.87), indicating substantial agreement. These findings support the reproducibility of expert morphology assessment in this dataset. However, because the analysis was conducted within a single institution, external interlaboratory reproducibility remains to be evaluated.

3.6. Evaluation of the Operational Dysmorphic RBC‐Based Morphology Scoring System

To assess sample‐level applicability, model‐predicted RBC classifications from the 82 validation images, each representing a distinct urine sample, were integrated into the operational dysmorphic RBC‐based morphology scoring system described above. Samples were stratified into Negative, Few, Moderate, and Major categories according to the percentage of dysmorphic RBCs among classifiable RBCs, as summarized in Table S1. These categories were used for feasibility assessment and were compared with manual expert assessment.

The concordance between AI‐based scoring and manual expert assessment is shown in Table 2. Among the 82 distinct validation samples, no samples fell into the Few category (15%–29% dysmorphic RBCs); therefore, performance in this subgroup could not be evaluated.

TABLE 2.

Concordance between AI‐based scoring and manual expert assessment across operational dysmorphic RBC‐based morphology categories in 82 distinct validation samples.

Clinical category (dysmorphic RBC score)	No. of samples	AI correct prediction	Concordance (%)	95% CI
Negative (< 15%)	39	39	100	91.0–100
Few (15%–29%)	0	N/A	N/A	N/A
Moderate (30%–59%)	11	9	81.8	48.2–97.7
Major (≥ 60%)	32	25	78.1	60.0–90.7

Note: Each validation sample was represented by one urine sediment image. Exact binomial 95% confidence intervals were calculated for sample‐level concordance rates. No validation samples were available in the Few category; therefore, performance in this subgroup could not be evaluated. The scoring categories were used as operational morphology‐based categories for feasibility assessment.

For samples in the Negative category (< 15% dysmorphic RBCs), the AI system achieved 100% concordance with expert assessment, with all 39 samples correctly classified (39/39; 95% CI, 91.0%–100%). For samples in the Moderate category (30%–59% dysmorphic RBCs), 9 of 11 samples were correctly classified, corresponding to a concordance rate of 81.8% (95% CI, 48.2%–97.7%). For samples in the Major category (≥ 60% dysmorphic RBCs), 25 of 32 samples were correctly classified, corresponding to a concordance rate of 78.1% (95% CI, 60.0%–90.7%).

These results indicate that the model performed best in identifying samples with low dysmorphic RBC proportions. However, the wide confidence intervals, particularly in the Moderate category, reflect the limited sample size of this pilot study. Therefore, these scoring results should be interpreted as preliminary evidence of morphology‐based concordance with expert assessment rather than externally validated diagnostic performance.

3.7. Efficiency of AI‐Assisted Interpretation

The computational inference time of the YOLOv5l model was compared with manual microscopic interpretation performed by 10 medical technologists. As shown in Table 3, the average manual interpretation time was 38.6 s per microscopic field, with individual averages ranging from 12.2 to 58.9 s depending on sample complexity and operator experience.

TABLE 3.

Comparison of manual microscopic interpretation time and AI computational inference time.

Method	Time measurement	Average time
Manual microscopy	Interpretation time per microscopic field under routine laboratory conditions	38.6 s
AI system (YOLOv5l)	Computational inference time per image	0.033 s

Note: AI inference time did not include image acquisition, image loading, preprocessing, expert review of AI‐generated outputs, or quality‐control procedures. Therefore, the timing analysis reflects computational inference speed rather than a complete end‐to‐end clinical workflow comparison.

In contrast, the average AI inference time was 0.033 s per image. This value represents computational inference time only and does not include image acquisition, image loading, preprocessing, review of AI‐generated outputs, or quality‐control procedures. Therefore, the timing results should not be interpreted as a complete end‐to‐end workflow comparison. Instead, they indicate that the trained model can perform rapid image‐level inference and may provide technical feasibility for preliminary urinary RBC morphology screening.

4. Discussion

In this retrospective pilot study, we developed and evaluated a YOLOv5l‐based deep learning system for urinary red blood cell (RBC) morphology classification using high‐resolution urine sediment images. The model classified RBCs into isomorphic, dysmorphic, and unknown categories and was further integrated with a dysmorphic RBC‐based scoring system for sample‐level morphology‐based stratification. The main findings were that the model showed moderate object‐level performance for dysmorphic RBC classification, high recall but low precision for isomorphic RBCs, preliminary concordance with expert assessment at the sample level, and rapid computational inference time. Importantly, the present study should be interpreted as a proof‐of‐concept feasibility study evaluating morphology classification and concordance with expert microscopy, rather than as a diagnostic accuracy study for biopsy‐confirmed glomerular disease.

A central issue in urinary RBC morphology assessment is that dysmorphic RBCs are biologically heterogeneous. As the early recognition that urinary RBC morphology can help distinguish glomerular from non‐glomerular bleeding, urine sediment microscopy has remained an important laboratory tool for evaluating hematuria [1]. However, dysmorphic RBCs are not a single morphological entity. They include acanthocytes, fragmented RBCs, ghost‐like cells, and other irregular forms, and these morphologies may differ in their specificity for glomerular bleeding [4, 6, 7]. Among them, acanthocytes, also referred to as G1 cells in some classification systems, are considered one of the most specific markers of glomerular hematuria because of their ring‐shaped appearance and vesicle‐like membrane protrusions [5, 6]. In the present pilot study, acanthocytic or G1‐like features were considered within the broader dysmorphic RBC category when identifiable, but acanthocytes were not annotated or analyzed as an independent subclass. This represents an important limitation because an acanthocyte‐specific model may provide greater biological specificity for future glomerular hematuria screening.

The present study addressed morphological uncertainty by using a three‐class annotation strategy. Instead of forcing every object into an isomorphic or dysmorphic RBC category, we included an “unknown” category for ambiguous objects, overlapping cells, poorly focused structures, ghost‐like cells, and background artifacts. This design reflects real‐world urine sediment microscopy, where not all observed objects can be confidently classified. The unknown category also provided a mechanism for retaining technically or biologically ambiguous objects during model development without assigning them to clinically meaningful RBC morphology classes.

The moderate recall for dysmorphic RBCs should be interpreted in the context of both biological heterogeneity and dataset imbalance. Dysmorphic RBCs are morphologically diverse and may include acanthocytic, fragmented, ghost‐like, doughnut‐like, and irregularly contoured forms. In addition, the annotated dataset contained substantially fewer dysmorphic RBCs than isomorphic RBCs. This imbalance likely reduced the diversity of dysmorphic features available for model learning and may have contributed to missed dysmorphic RBCs, particularly in cells with subtle membrane protrusions, partial borders, or borderline morphology. Therefore, the current dysmorphic RBC performance should be regarded as preliminary. Future studies should increase the number and diversity of dysmorphic RBC images and evaluate class‐balancing approaches, such as targeted sampling, oversampling, class‐weighted loss functions, focal loss, and augmentation strategies focused on acanthocytic and other clinically informative dysmorphic forms.

The high recall for isomorphic RBCs indicates that the model was effective in detecting RBCs with relatively preserved morphology. However, the low precision of 0.47 for isomorphic RBCs indicates that a substantial proportion of predicted isomorphic RBCs were false positives. Qualitative review suggested that these false positives were mainly associated with background artifacts, air bubbles, poorly focused objects, ghost‐like cells, overlapping RBCs, and partially preserved circular structures that resembled isomorphic RBCs. This error pattern is important because false‐positive RBC detection may affect downstream sample‐level dysmorphic RBC scoring by altering the denominator used for RBC proportion calculations. Therefore, the current model should not be used as an autonomous reporting system. Qualitative review of representative validation examples (Figure 3) further suggested that false‐positive or uncertain predictions were mainly associated with background artifacts, ghost‐like structures, and overlapping cell boundaries. Future versions should incorporate a broader range of artifact and non‐RBC objects during training, consider a dedicated background or artifact class, and include quality‐control rules to flag images with excessive overlap, poor focus, or ambiguous RBC‐like structures.

Urine physicochemical conditions are important confounders in urinary RBC morphology assessment. Urine pH, specific gravity, and osmolality can alter RBC appearance through swelling, hemolysis, ghost‐cell formation, shrinkage, or crenation, thereby affecting both manual microscopy and image‐based classification [7, 9]. Although the present annotation strategy assigned ghost‐like and severely distorted objects to the unknown category when reliable classification was not possible, urine pH, specific gravity, and osmolality were not systematically incorporated into model development or subgroup analysis. Therefore, the present results cannot determine whether model performance is stable across different urine physicochemical conditions. Future prospective studies should record these variables and evaluate their effects on both expert annotation and AI‐based classification.

The relationship between image‐based deep learning and automated urinary particle analyzers also deserves careful interpretation. Modern automated urine analyzers, including flow cytometry‐based systems such as the Sysmex UF series, have improved the throughput and standardization of routine urinalysis [10]. Parameters such as urinary RBC distribution and small RBC‐related measurements may provide useful information for differentiating glomerular from non‐glomerular hematuria [11, 12]. However, these systems generally rely on indirect physical properties, such as cell size and scatter characteristics, rather than direct visual recognition of membrane protrusions or acanthocytic morphology. The present image‐based approach is therefore best considered complementary to existing technologies. Because paired UF‐5000 or other urinary flow cytometry‐derived parameters were not available in this retrospective image dataset, a direct head‐to‐head comparison could not be performed. Such a comparison will be necessary to determine whether morphology‐aware deep learning provides incremental value beyond existing automated urinalysis platforms.

The sample‐level scoring analysis showed 100% concordance with expert assessment in the Negative category, defined as < 15% dysmorphic RBCs. This finding suggests that the model may help identify samples with low dysmorphic RBC proportions in a morphology‐based screening context. However, this result should be interpreted cautiously. First, the confidence interval around this estimate reflects the limited sample size of this pilot study. Second, no validation samples fell into the Few category, preventing evaluation of this clinically relevant intermediate subgroup. Third, the scoring categories used in this study were operational morphology‐based categories for feasibility assessment and were not externally validated diagnostic thresholds. Therefore, the present scoring analysis should be viewed as expert‐concordance assessment rather than evidence of definitive clinical diagnostic performance.

The potential clinical relevance of morphology‐aware urinary RBC screening is particularly important in immune‐mediated glomerular diseases. IgA nephropathy, lupus nephritis, and ANCA‐associated vasculitis commonly involve glomerular inflammation or capillary wall injury and may present with microscopic or macroscopic hematuria. In these conditions, recognition of dysmorphic RBCs may provide an early laboratory clue suggesting glomerular involvement and may help prioritize nephrological evaluation. However, urinary RBC morphology alone cannot distinguish among specific immune‐mediated kidney diseases and should not be interpreted independently of proteinuria, kidney function, urinary casts, serological markers, clinical presentation, and, when indicated, renal biopsy. Therefore, the proposed system may be most appropriately positioned as a preliminary morphology‐based workflow support tool for identifying samples that warrant expert review, rather than as a disease‐specific diagnostic system.

A major limitation of this study is the absence of an independent clinical reference standard. The present analysis evaluated AI‐based morphology classification against expert microscopic assessment, but did not validate performance against biopsy‐confirmed glomerular disease, nephrologist adjudication, or longitudinal clinical outcomes. Therefore, this study cannot establish diagnostic accuracy for glomerular hematuria or specific kidney diseases. Urinary RBC morphology should be interpreted together with clinical history, proteinuria, kidney function, serological markers, imaging findings, and, when clinically indicated, renal biopsy. Future studies incorporating biopsy‐confirmed diagnoses or nephrologist‐adjudicated clinical endpoints are needed to determine the clinical diagnostic value of AI‐assisted urinary RBC morphology analysis.

Generalizability is another important limitation. This was a single‐center retrospective study using images acquired under a standardized local sample‐preparation protocol and a specific microscope‐camera system. Although the dataset was partitioned at the urine‐sample level to reduce data leakage between training and validation sets, no independent external test cohort was available. Therefore, the robustness of the model across different laboratories, microscope platforms, imaging settings, specimen preparation procedures, and patient populations remains unknown. Multicenter external validation using independent image datasets is required before broader clinical implementation. In addition, each validation sample was represented by a single urine sediment image, which may not fully capture within‐sample morphological heterogeneity across multiple microscopic fields.

Several additional methodological limitations should also be noted. First, the dataset was imbalanced, with substantially more isomorphic than dysmorphic RBCs, which may have affected dysmorphic RBC recall. In addition, background‐related false‐positive isomorphic RBC predictions may have influenced downstream sample‐level scoring by altering the total classifiable RBC count. Second, ROC/AUC analysis could not be performed because complete object‐level prediction confidence scores were not retained in the archived retrospective object‐detection outputs. Future prospective analyses should retain object‐level confidence scores to enable ROC, precision–recall curve, and threshold‐optimization analyses. Third, acanthocytic or G1‐like RBCs were not analyzed as an independent subclass. Fourth, urine pH, specific gravity, and osmolality were not systematically incorporated into subgroup analysis. Fifth, paired UF‐5000 or other automated urinary particle analyzer data were not available. Sixth, performance in the Few category could not be evaluated because no validation samples fell into this range. Seventh, although interobserver agreement was assessed at both the sample level and in a stratified object‐level subset, the reproducibility analysis was performed within a single institution, and full object‐level re‐annotation of the entire dataset was not conducted. Finally, AI inference time reflected computational processing only and not the full clinical workflow.

The efficiency results also require cautious interpretation. The reported AI processing time represents computational inference time per image only and does not include image acquisition, image loading, preprocessing, review of AI‐generated outputs, or laboratory quality‐control procedures. In contrast, manual microscopy time reflects human interpretation under routine laboratory conditions. Therefore, the present timing analysis should not be interpreted as a complete end‐to‐end workflow comparison. Rather, it demonstrates that once images are available, the trained model can perform rapid computational inference, supporting the technical feasibility of future workflow integration.

Quality assurance remains essential for urinary sediment microscopy and for any AI‐assisted implementation. Annotation reproducibility is particularly important for urinary RBC morphology analysis because dysmorphic RBCs include borderline and heterogeneous forms. In the present study, two complementary interobserver agreement analyses were performed. At the sample level, independent review of the 82 validation images showed high agreement for dysmorphic RBC‐based morphology scoring, with a Cohen's kappa of 0.88 and a weighted Cohen's kappa of 0.95. At the object level, independent re‐annotation of a stratified random subset of 500 RBC or RBC‐like objects showed substantial agreement, with a Cohen's kappa of 0.82. These findings support the reliability of the expert morphology assessment used in this pilot dataset. Nevertheless, external quality assessment programs have shown that urinary sediment particle identification and clinical interpretation can vary among laboratories, and structured quality schemes may improve both technical recognition and clinical understanding [19]. Future studies should incorporate blinded multi‐observer annotation, acanthocyte‐specific labeling, external quality assessment materials, and multicenter proficiency datasets to further strengthen annotation reliability and interlaboratory reproducibility.

The use of deep learning in urinary sediment analysis is rapidly evolving. Recent studies have moved beyond simple image classification toward object detection, instance segmentation, prospective clinical validation, and specialized architectures for small or morphologically diverse urinary particles [20, 21, 22, 23, 24]. For example, YOLO‐based instance segmentation has been applied to urine formed elements to improve the detection and segmentation of small and dense urinary particles. More recent urinary sediment studies have also explored transformer‐based or multi‐head object‐detection architectures, self‐supervised pretraining, and prospective laboratory integration to address challenges such as data scarcity, class imbalance, image variability, and generalizability. These developments support the broader feasibility of AI‐assisted urinalysis, but they also highlight the need for external validation, confidence‐score retention, interpretability analysis, and prospective workflow evaluation before clinical implementation. In contrast to general urinary particle classification, the present study focuses specifically on urinary RBC morphology and the distinction among isomorphic, dysmorphic, and unknown categories, which requires careful handling of biologically ambiguous forms and artifact‐like objects.

In conclusion, this retrospective pilot study demonstrates the feasibility of a morphology‐aware YOLOv5l‐based system for urinary RBC classification in urine sediment images. By incorporating isomorphic, dysmorphic, and unknown categories and integrating model outputs with an operational dysmorphic RBC‐based scoring framework, the proposed approach showed preliminary concordance with expert microscopic assessment and rapid computational inference. However, these findings should be interpreted as evidence of morphology‐classification feasibility and workflow‐support potential, not as standalone diagnostic performance for glomerular hematuria or biopsy‐confirmed kidney disease. Multicenter external validation using independent test cohorts, clinical reference standards, urine physicochemical variables, acanthocyte‐specific annotation, comparison with automated urinary particle analyzers, expanded multi‐observer reproducibility assessment, and external quality assessment materials will be required before clinical implementation.

Author Contributions

Yih‐Lon Lin contributed to deep learning model development, algorithm implementation, and data analysis. Jung‐Sheng Chen contributed to methodological consultation, data interpretation, and manuscript review and editing. Ya‐Fan Chuang contributed to urine sediment image collection, morphological annotation, and laboratory interpretation. Siang‐Ru Huang contributed to data organization, literature review, and manuscript preparation. Chien‐Sen Liao contributed to study conception and design, project supervision, manuscript writing, critical revision, and final approval. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the National Science and Technology Council, Taiwan (grant nos. NSTC 115‐2221‐E‐214‐012 and NSTC 115‐2515‐S‐214‐004), and I‐Shou University/E‐Da Hospital (grant nos. ISU‐115‐IUC‐04 and EDAHJ113005). The funders had no role in the study design; data collection, analysis, or interpretation; manuscript preparation; or the decision to submit the manuscript for publication.

Ethics Statement

This study was approved by the Institutional Review Board of Pingtung Christian Hospital (IRB No. IRB731A). The requirement for informed consent was waived because this retrospective study used de‐identified images and laboratory data.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Operational conversion of dysmorphic RBC scoring categories into percentage‐based thresholds for AI‐assisted analysis.

Table S2: Interobserver agreement for dysmorphic RBC morphology assessment.

Data Availability Statement

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request.