Deep learning based instance segmentation of mandarin fruit slices for precision assessment and morphological quantification

Abstract

Accurate instance segmentation of mandarin fruit slices is essential for quantifying segment morphology and central core structure, which are key traits in cultivar evaluation, fruit quality assessment, and postharvest application. Manual measurement of these anatomical features, however, is time-consuming and prone to inconsistency. In this study, we developed an automated segmentation framework based on YOLOv8 to detect and semantic segment the fruit segments and the central core in high-resolution mandarin transversely cut images. The model was trained on a curated datasetderived from 58 original high-resolution cross-sectional images (5100 × 7019 pixels), which were systematically partitioned into 280 cropped sub-images (1712 × 1778 pixels), each containing a single complete citrus slice, and demonstrated excellent performance. YOLOv8 achieved near-perfect detection metrics, with bounding box metrics precision ~ 0.997, recall ~ 1.00, mAP50 ~ 0.995, and mAP50-95 ~ 0.922. Semantic segmentation accuracy was similarly strong, with precision = 0.997, Recall ~ 1.00, mAP50 ~ 0.995, and mAP50-95 = 0.965. Training and validation losses converged steadily, indicating stable learning without overfitting. The nonsignificant differences between YOLOv8-predicted measurements and ground truth data, together with low mean absolute error (MAE) values, demonstrate that the model not only performs well in semantic segmentation metrics but also maintains high accuracy in quantitative measurements, which is critical for cultivar discrimination and genetic studies. Our work provides a robust, high-precision, and reproducible framework for mandarin fruit slice phenotyping, offering significant potential for applications in agricultural research, breeding programs, and automated fruit quality evaluation.

Introduction

Over the past decade, global consumption of fresh, easy-to-peel mandarins has increased steadily^1–3. Despite their growing commercial importance, scientific knowledge of mandarin fruit traits remains limited compared with that of other citrus species, particularly oranges, which have been the primary focus of citrus research^3,4. As consumer preferences continue to shift toward mandarins, there is an increasing need for systematic characterization of their distinctive fruit quality traits⁵.

Mandarin is among the most genetically and phenotypically heterogeneous groups within the genus Citrus, exhibiting wide variation in both morphological and anatomical traits^6,7. Although most citrus species are predominantly self-pollinated, even low levels of cross-pollination can generate zygotic seedlings, further increasing phenotypic variability^8,9. Consequently, mandarin fruit exhibits a complex internal anatomy, and comprehensive phenotypic characterization of these traits is essential for breeding, cultivar evaluation, and fruit quality improvement^3,10. Accurate morphological and anatomical characterization is therefore a fundamental requirement for assessing genetic diversity and for selecting superior genotypes with desirable attributes such as fruit size and shape, ease of peeling, seed content, and internal structure^7,11,12. However, traditional approaches to fruit anatomical analysis rely heavily on manual measurements, which are labor-intensive, time-consuming, and poorly suited to large-scale phenotyping^13–15.

High-throughput phenotyping and plant phenomics offer powerful solutions to these limitations by enabling rapid, objective, and non-destructive measurement of complex traits across large populations^13,16,17. Advances in deep learning-based computer vision, particularly convolutional neural networks (CNNs), have transformed fruit phenotyping by enabling automated detection, semantic segmentation, and quantitative trait extraction^15,17. Such approaches have been successfully applied to fruit detection, yield estimation, texture analysis, and quality assessment across multiple crops, including apple, tomato, and grape…^18,19. Within this context, the YOLOv8 framework provides a unified architecture for object detection and instance segmentation, combining high accuracy with computational efficiency and making it well suited for scalable phenomics pipelines^20–23.

In citrus research, recent studies have demonstrated the potential of deep learning for fruit detection, maturity or ripeness classification, and disease or decay identification using RGB images as well as X-ray or hyperspectral data^18,24–27. These studies have contributed to advances in citrus production, particularly in automated grading and sorting, as well as disease monitoring and management^16,28–31. However, most existing research has focused primarily on external fruit traits, such as color, size, and shape, which represent only a subset of the phenotypic information relevant to fruit development and quality^20–22,32. Despite the biological importance of internal mandarin fruit anatomical traits including fruit segment number, fruit segment area, and central core structure, which are closely linked to developmental processes such as ovary locule formation, vascular development, and tissue differentiation, no published studies have specifically focused on quantitative parameter estimation from transversely cut mandarin fruits^14,33. Therefore, precise instance segmentation of sliced citrus fruit, particularly for extracting internal anatomical traits such as fruit segments and the central core, remains poorly documented^13,15.

This study addresses a critical gap in citrus plant phenomics by proposing an automated, YOLOv8-based pipeline for the detection, semantic segmentation, and quantitative analysis of internal anatomical traits from transversely cut mandarin fruit images. By integrating accurate instance segmentation with biologically meaningful trait extraction, this approach enables scalable and reproducible phenotyping of mandarin fruit anatomy. The proposed framework provides a foundation for large-scale genetic studies, cultivar evaluation, and breeding programs, supporting data-driven selection of mandarin genotypes with desirable morphological and anatomical characteristics.

Materials and methods

The proposed system is presented in Fig. 1. First, high-resolution images of citrus cross-sections are captured. The images are then annotated using the Labelme software. After annotation, the original images are cropped to obtain individual citrus slices, where each image contains a single complete cross-section. The resulting dataset is subsequently divided into training, validation, and test sets with a ratio of 70:15:15. The training set is used to train the YOLO instance segmentation model to obtain a trained model. The trained model is then applied to predict citrus slices, including partially cut samples, and generate segmentation masks. These masks are subsequently used to compute morphological parameters of the citrus slices.

Fig. 1.

Block diagram of the proposed citrus phenotyping system. The workflow includes image acquisition, annotation using Labelme, cropping of individual citrus slices, dataset splitting into training, validation, and test sets, training of the YOLO instance segmentation model, and extraction of morphological parameters from the predicted segmentation masks.

Image acquisition and preprocessing

In the task of precise semantic segmentation of the peel, fruit segments, and central core of mandarin fruits, the acquisition and preprocessing of plant images constitute the foundational component of this study. Fresh mature mandarin fruits were harvested from the experimental field of the Institute of Sustainable and Environmentally Friendly Agricultural Sciences following a standardized protocol. The fruits were transversely cut along their diameter and meticulously scanned using an Epson color scanner equipped with an optical contact image sensor (Epson Perfection V39, Japan), generating high-resolution images in JPEG format at 5100 × 7019 pixels and 600 dpi (dots per inch). A black background was used to enhance contrast with the fruit tissue, and custom 3D panels were employed to ensure accurate representation of the fruits’ actual size (Fig. 2, upper), where the displayed slices represent different mandarin fruits rather than multiple sections from a single fruit.

Fig. 2.

Image preprocessing results for cross-sectional slices obtained from different mandarin fruits. The figure illustrates the steps used to prepare the dataset, including extraction of individual citrus slices from the original scanned images and preparation of images used for annotation and model training.

The preprocessing procedure consisted of two steps: (1) renaming the image files and (2) partitioning each original mandarin slice image into non-overlapping sub-images. First, the original filenames were replaced using a standardized naming convention, {N}{Image number}.jpg, where N denotes the mandarin cultivar. Subsequently, a total of 58 original high-resolution images were acquired. Each image contained multiple mandarin cross-sections and was divided into non-overlapping sub-images. From this process, 280 cropped sub-images (1712 × 1778 pixels) were obtained, each containing a single complete citrus cross-section. These sub-images were directly used for annotation and model development.. This partitioning improved the clarity of structural features, such as the peel, fruit segments, and central core, and reduced the computational load for downstream deep learning tasks, given the high resolution of the original images. This approach substantially lowered memory requirements and accelerated both model training and inference. The processed sub-images were saved using a new filename format, {N}{Image number}{XY}.jpg, where XY denotes the grid index (Fig. 2, lower).

Annotation for instance segmentation

To facilitate instance segmentation of transversely cut mandarin fruits, LabelMe software was employed as a graphical image annotation tool to manually draw polygons around fruit segments section and the central core, assigning the corresponding labels³⁴. The resulting annotations were exported in JSON format, enabling efficient analysis and quantification of mandarin cross-sectional morphology. During preprocessing, the dataset was divided into training, validation, and testing subsets with a ratio of 70:15:15 to ensure balanced data representation and reliable model evaluation. The training set (70%) was used to learn the model parameters, while the validation set (15%) was employed during training to monitor performance and adjust model settings, helping to reduce overfitting. The remaining 15% of the data was reserved as an independent test set for final performance evaluation, providing an unbiased assessment of the model’s robustness and generalization capability.

Model architecture and operating environment

The semantic segmentation of the fruit segments section and central core within a mandarin fruit sliced image was performed using a deep learning based object detection and semantic segmentation framework developed with the YOLOv8 architecture³⁵. YOLOv8 was selected due to its optimized balance of speed and accuracy, making it suitable for high-throughput phenotyping tasks involving complex fruit anatomy. The dataset used for model development consisted of high-resolution images of mandarin fruit slices, each manually annotated to delineate key anatomical structures, including the fruit segment tissues and central core. All images were formatted into YOLO-compatible semantic segmentation labels to support pixel-level feature extraction.

The high-resolution images of mandarin fruit slices exhibited substantial variability in fruit morphology, illumination, slice thickness, and tissue coloration, enabling the model to learn from diverse phenotypic presentations. This variability was further enhanced through data augmentation procedures including random affine transformations, horizontal flipping, rotation, and brightness-contrast adjustments which have been widely shown to improve model generalization under non-uniform imaging conditions. Model training was conducted using a YOLOv8 pretrained backbone to exploit transfer learning advantages, thereby accelerating convergence and improving feature extraction efficiency for fine-scale fruit structures. The network was optimized using a composite loss function comprising box regression loss, semantic segmentation mask loss, classification loss, and distribution focal loss. These metrics were monitored throughout training to ensure steady improvement in boundary detection and class differentiation. The model was trained for 100 epochs, with batch size and learning rate parameters tuned based on GPU memory availability and convergence behavior observed during early training iterations.

All training and inference experiments were executed on a high-performance computational platform equipped with Windows 11 Professional (64-bit) and powered by an AMD Ryzen 7 5800X3D 8-core processor and 32 GB RAM. Model training utilized an NVIDIA GeForce RTX 4080 GPU with 16 GB of dedicated memory, enabling accelerated tensor computations and efficient handling of large image batches. This hardware configuration provided the computational throughput necessary for processing thousands of high-resolution fruit slice images and supported rapid experimentation during model refinement.

Model evaluation and performance metrics

A meticulous evaluation of model performance against ground truth annotations is essential for validating the accuracy and reliability of fruit-slice semantic segmentation algorithms. The ground truth, representing manually annotated regions of the fruit segments region and central core, serves as the reference for assessing the predictions generated by YOLOv8. These annotations, created through careful manual delineation, define the precise locations and boundaries of key fruit structures, such as fruit segments and central core, providing a human-verified standard for benchmarking automated detection and semantic segmentation methods. YOLOv8, a state-of-the-art object detection framework, predicts these structural components using bounding boxes accompanied by confidence scores, with semantic segmentation applied within each detected region. Renowned for its efficiency in real-time image analysis, YOLOv8 produces rapid and accurate predictions. The model outputs are evaluated against the ground truth annotations using key performance metrics, including precision, recall, and Intersection over Union (IoU), which collectively provide a quantitative measure of detection and semantic segmentation accuracy. This evaluation enables a comprehensive assessment of YOLOv8’s ability to capture the intricate morphological features of mandarin slice fruits. Metrics such as IoU, precision, recall, and F1 score quantify the alignment between predicted fruit regions and ground truth annotations, offering insights into the strengths and limitations of the algorithm^36,37.

Intersection over Union (IoU), also known as the Jaccard Index, measures the overlap between the predicted semantic segmentation mask and the ground truth mask (1), where GTBbox is the ground truth bounding box and PBbox is the predicted bounding box. IoU values ≥ 0.5 were considered successful detections (true positives, TP), with higher values indicating more precise delineation of object boundaries. Precision quantifies the proportion of correctly predicted positive instances relative to all positive predictions (2), while Recall measures the proportion of correctly predicted positive instances relative to all actual positives (3). The F1 Score (4), defined as the harmonic mean of precision and recall, provides a balanced measure of model performance, particularly for datasets with class imbalance. Average Precision (AP) was computed for each class independently, and the mean Average Precision (mAP) was calculated to represent overall performance across all classes. Higher AP and mAP values indicate superior detection and semantic segmentation accuracy. Such rigorous evaluation not only validates the effectiveness of YOLOv8 for mandarin fruit-slice semantic segmentation but also guides further refinement for improved precision and reliability in complex image analysis tasks.

1

2

3

4

In addition to quantitative metrics, visual inspection of predicted masks overlaid on original images provided qualitative insights into model performance. This approach allowed identification of discrepancies between predicted and actual boundaries, evaluation of semantic segmentation smoothness, and detection of false positives or false negatives. Comparative visualization of YOLOv8 outputs alongside ground truth enabled a nuanced understanding of the model’s strengths and limitations under various conditions, such as overlapping fruit segments or variable lighting. Confidence scores generated by YOLOv8 for each detected region further assisted in evaluating detection reliability. High-confidence mispredictions may indicate overfitting or insufficient training, whereas low-confidence correct detections suggest areas for model refinement. By integrating both quantitative and qualitative analyses, a comprehensive understanding of YOLOv8’s performance in mandarin fruit semantic segmentation was achieved, providing a basis for further optimization and deployment in precision phenotyping studies.

Parameter computation

Image-based phenotyping of citrus fruit was performed using a custom Python pipeline designed to extract anatomical and morphological parameters from YOLO instance segmentation outputs. The workflow consisted of four stages: (i) loading semantic segmentation files generated by YOLO, (ii) polygon reconstruction and ellipse fitting, (iii) extraction of geometric and anatomical traits, and (iv) data export to a structured CSV file.

For each image, polygon annotations were extracted and converted to pixel coordinates. Ellipse fitting was performed using the least-squares method implemented in OpenCV (cv2.fitEllipse). This produced the geometric parameters of each anatomical structure, including ellipse center, major and minor axes, and rotation angle. Polygon areas were calculated using the shoelace formula, and the largest polygon for each object class was retained as the representative central core or fruit segment boundary.

To isolate the full fruit morphology, an expanded version of the segment ellipse was used to generate a binary fruit mask. The masked image was converted to grayscale, smoothed using a Gaussian filter, and binarized using Otsu thresholding. Morphological closing and opening operations removed noise and small artifacts. The largest connected contour was extracted from the cleaned binary mask and used to fit a final fruit ellipse, from which fruit width (minor axis) and fruit length (major axis) were derived in pixel units.

Based on the extracted contour points, multiple geometric and morphological traits were computed, including the area of the total fruit cross secctional (A_fruit), the area of the fruit segmented region (A_segments), and the central core (A_central). The peel area was computed as:

For peel thickness estimation, a radial distance-based approach was applied. The ellipses were fitted to both the outer fruit contour and the fruit segment boundary. For each boundary point, the Euclidean distance from the ellipse center to the outer fruit contour (R_fruit) and to the inner fruit segments contour (R _segments) was calculated. Local peel thickness was defined as:

where denotes the angular position around the fruit center. Mean peel thickness was obtained by averaging across all sampled angles.

The workflow used to determine the number of citrus fruit segments from the segmented images is illustrated in Fig. 3. First, the cropped citrus slice image is processed by the YOLO instance segmentation model to obtain the mask of the central core. The contour points of the core mask are then extracted, and an ellipse is fitted to these points to estimate the geometric center of the core. This center point is used as a reference to calculate the radial distance between the center and each boundary point of the fruit contour, generating a radial distance profile. To estimate the number of fruit segments, peaks in this radial distance signal are detected using a prominence-based algorithm implemented with scipy.signal.find_peaks. Each detected peak corresponds to the outermost position of an individual fruit segment, and the total number of peaks is recorded as the estimated number of anatomical fruit segments, as illustrated in Fig. 4.

Fig. 3.

Flow diagram illustrating the process of fruit segment counting based on YOLO instance segmentation, contour extraction, ellipse fitting, and radial distance peak detection.

Fig. 4.

Visualization of detected peaks used to estimate the number of anatomical fruit segments in the output image. The blue points show the detected peaks from fruit segments contour points (far the center of the images) and from central core contour points (near the center of the images).

Finally, the extracted parameters were consolidated into a structured data record containing: image filename and dimensions, central core area, fruit segment area, fruit area, number of fruit segments, fruit width and fruit length for each image (Table 1). All measurements were written to a comma-separated values (CSV) file using Python’s built-in csv module. Individual visualizations of cleaned fruit masks and peak-detected segment landmarks were saved for quality control.

Table 1.

Morphometric traits of mandarin sliced fruit.

Trait	Description	Importance in fruits morphological traits
Fruit width (pixel)	The maximum horizontal distance across the fruit cross-section, measured perpendicular to the longitudinal axis	This trait reflects fruit girth and serves as a primary indicator of fruit size, commonly used for grading and validation of quantitative fruit characteristics38,39
Fruit length (pixel)	The maximum distance along the longitudinal axis of the fruit cross-section	This trait represents fruit elongation. Together with fruit width, it is a key indicator of overall fruit size and shape38,39
Fruit area (pixel2)	The total two-dimensional area enclosed by the outer fruit boundary in the cross-sectional image	This parameter serves as an integrated descriptor of overall fruit size and biomass allocation38
Fruit peel area (pixel2)	The area between the outer fruit boundary and the inner boundary separating peel from fruit segments	It reflects rind thickness and structural investment, which are important for fruit protection, storability, and transport resistance21,25
Peel thickness (pixel)	The average radial distance between the outer fruit boundary and the inner peel-segment boundary, measured across the fruit circumference	It provides an estimate of rind thickness, including the epicarp (flavedo) and mesocarp (albedo), which contain high concentrations of essential oils and are relevant to fruit protection, handling quality, and shelf life12,27
Fuit segment area (pixel2)	The cumulative area of all individual fruit segments within a slice corresponds to the edible portion of the fruit	This edible region, contains seeds and juice vesicles connected to the central core and directly influences juice yield and eating quality2,4
Fruit segment number	The total count of distinct fruit segments identified within a single cross-section	This trait is genetically regulated and is commonly used for cultivar identification and evaluation4,12
Fuit central core area (pixel2)	The area of the central pith or vascular core region located at the geometric center of the fruit slice	This trait characterizes internal anatomical organization and tissue differentiation, which vary among citrus species1,4,5

Statistical analysis

Semantic segmentation performance of the YOLOv8 model was quantitatively evaluated during training and validation using standard metrics, including intersection over union (IoU), precision, recall, and F1-score. Following validation of semantic segmentation accuracy, morphological traits were extracted from the predicted masks of the fruit segments and central core. The evaluated parameters included fruit width, fruit length, total fruit area, segment area, number of fruit segments per fruit, and individual segment areas for each sliced mandarin fruit. All extracted datasets were first assessed for normality using the Shapiro–Wilk test. As all traits exhibited normal distributions, parametric statistical analyses were subsequently applied. Differences between ground truth measurements and YOLOv8-predicted values were assessed using a two-sample t-test. In addition, prediction accuracy was quantified by calculating the mean absolute error (MAE) for each morphological parameter. Data visualization was performed using boxplots to illustrate the distribution and agreement between ground truth measured and predicted values. All statistical analyses and graphical visualizations were conducted in R software (version 4.5.1; R Foundation for Statistical Computing, Vienna, Austria) using the ggplot2 package.

Results

In this study, the mandarin slice dataset was annotated using the YOLOv8 instance segmentation format. It was partitioned into training, validation and testing subsets at a ratio of 70:15:15. The final dataset consisted of 280 mandarin cross-section images, comprising 196 training images, 42 validation images and 42 testing images. The training subset contained 196 annotated instances corresponding to the fruit segments and central core, each exhibiting notable anatomical variability. The validation and testing subset included 42 instances with comparable morphological distributions.

Comprehensive statistical summaries of the dataset are presented in Fig. 5. The upper-left panel shows the absolute frequency of each anatomical class, while the upper-right panel illustrates the distribution of bounding-box areas generated from segmentation masks. The lower-left panel displays the spatial occurrence of annotated structures, indicating no positional bias across the dataset. The lower-right panel details size characteristics, including width and height distributions for each class. These analyses confirm that anatomical features are uniformly and randomly distributed across mandarin slice images. Relative instance widths range from 0.08 to 0.85 of the total image width, whereas heights range from 0.15 to 0.85 of the image height. Collectively, this information confirms the robustness and representativeness of the dataset for training and evaluating YOLOv8-based segmentation models for mandarin fruit morphology analysis.

Fig. 5.

Summary of annotations in the mandarin slice dataset. The figure shows the distribution and examples of annotated classes, including the central core and fruit segments, used for training the YOLO instance segmentation model.

To analyze the structural characteristics of mandarin fruit slices, accurate detection and segmentation of anatomical regions serve as the essential step in the image analysis pipeline. The YOLOv8 instance segmentation model achieved high accuracy in detecting and segmenting fruit segments and the central core in sliced mandarin images. Quantitative evaluation using the Intersection over Union (IoU) metric showed substantial overlap between predicted masks and ground-truth annotations. The mean IoU exceeded 95% for fruit segments and was over 80% for the central core, indicating precise delineation of segment regions with well-defined boundaries. IoU is a widely used metric for evaluating instance segmentation performance, as it directly quantifies the degree of overlap between predicted and true regions, with higher values indicating better model accuracy. The accompanying confidence scores (e.g., 0.98) reflect the model’s certainty in correctly identifying each region (Fig. 6), with higher values indicating stronger predictive confidence. The high IoU values obtained across validation samples underscore the robustness and reliability of the YOLOv8 model for instance segmentation of mandarin fruit slices.

Fig. 6.

Visualize of instance segmentation on output image. The predicted bounding boxes and segmentation masks for the central core and fruit segments are overlaid on the original images to illustrate the model’s detection and segmentation performance.

Following training, the optimized YOLOv8 model was deployed for automated analysis of sliced mandarin fruit images, enabling reliable segmentation of fruit segments and the central core. Figure 6 presents a detailed overview of the training and validation log archive for the YOLOv8 model applied to the sliced mandarin fruit dataset. The model converged successfully after 100 epochs, achieving low training losses, including a box loss of approximately 0.27, segmentation loss of 0.28, classification loss of 0.23, and distribution focal loss (DFL) of 0.88. These loss values indicate effective optimization of model predictions, reflecting accurate bounding box localization, precise delineation of fruit segments and the central core, reliable classification of target regions, and stable confidence distribution throughout training. For bounding box detection of fruit segments and the central core, the model achieved very high precision (0.994) and perfect recall (1.000), with an mAP@0.5 of 0.995 and an mAP@0.5–0.95 of 0.913. These results indicate highly accurate and consistent localization of the target anatomical regions across the validation dataset. Similarly, for mask segmentation, the model achieved a precision of 0.994 and perfect recall (1.000), with an mAP@0.5 of 0.995 and an mAP@0.5–0.95 of 0.966, demonstrating highly accurate and reliable delineation of fruit segments and the central core. The final training losses further demonstrate the model’s stability and strong generalization capability, with box loss converging to approximately 0.27, segmentation loss to 0.28–0.30, classification loss to about 0.23, and distribution focal loss (DFL) to approximately 0.88. These consistently low loss values at the final epoch indicate stable optimization and effective convergence of the network (Fig. 7).

Fig. 7.

Model training performance over 100 epochs. This figure shows the training log of the YOLOv8 model across 100 epochs, including metrics such as training and validation loss, segmentation accuracy, and mean average precision (mAP). These plots provide insights into the model’s convergence behavior, performance consistency, and generalization ability throughout the training process.

The performance curves of the YOLOv8 model, including F1-score, precision, and recall for both bounding box detection and mask segmentation across varying confidence thresholds, as well as precision-recall curves, was shows in Fig. 8. The precision and recall curves demonstrate how prediction accuracy and detection completeness vary with increasing confidence thresholds, respectively, reflecting the model’s ability to reliably identify fruit segments and the central core. The precision-recall (PR) curve further highlights the trade-off between precision and recall under different thresholds, which is particularly informative for evaluating segmentation performance across potentially imbalanced classes. The F1-score curve summarizes this relationship by combining precision and recall into a single metric, serving as an overall indicator of model performance. As the harmonic mean of precision and recall, a high F1-score indicates that the model maintains both low false-positive and low false-negative rates. Collectively, these curves confirm the robustness and reliability of the YOLOv8 model for accurate bounding box localization and mask segmentation, and demonstrate how threshold selection influences the prediction of quantitative fruit characteristics using image-based analysis.

Fig. 8.

Performance evaluation curves of YOLOv8 model for bounding box and mask predictions. This figure presents the detailed performance metrics of the trained YOLOv8 model. Panels (A–D) show the results for bounding box predictions: (A) F1 score curve, (B) precision curve, (C) precision-recall curve, and (D) recall curve. Panels (E–H) depict the corresponding metrics for instance mask predictions: (E) F1 score curve, (F) precision curve, (G) precision-recall curve, and (H) recall curve. These curves illustrate the model’s detection and segmentation performance across varying confidence thresholds, helping to assess its overall effectiveness and reliability in identifying and segmenting fruits characteristics.

In addition, the model achieved an average precision of approximately 0.997, perfect recall (1.000), and an F1-score of approximately 0.998, further confirming its excellent detection and segmentation performance.. These results indicate outstanding accuracy and robustness in detecting and segmenting anatomical structures in the sliced mandarin fruits dataset. The high precision indicates minimal false positives, while the perfect recall confirms that all annotated fruit segments and central cores were successfully detected by the model. The resulting F1-score reflects a strong balance between precision and recall, underscoring the reliability of the model. Overall, the consistently high IoU, precision, recall, and F1-score values confirm the robustness and effectiveness of YOLOv8 for accurate detection and instance segmentation of fruit segments and the central core in mandarin fruit characteristics.

The YOLOv8 model accurately predicted all instances of fruit segments and the central core with high confidence, demonstrating strong agreement with the ground truth annotations (Fig. 9). Using polygon reconstruction and ellipse fitting, several geometric and anatomical traits of sliced mandarin fruit were extracted, including fruit width, fruit length, fruit area, number of segments, segment area, and central core area. These traits are of significant importance for fruit breeding, cultivar evaluation, and the validation of quantitative fruit characteristics using image-based analysis.

Fig. 9.

Qualitative evaluation of instance segmentation by the YOLOv8 model on the dataset. Example images from the dataset are shown with predicted segmentation masks and bounding boxes, highlighting the model’s ability to accurately identify the central core and individual fruit segments.

Based on the YOLOv8-predicted segmentation results and the ground truth data, key morphological parameters were calculated and systematically compared. The mean absolute error (MAE) between predicted and ground truth values was computed to quantify prediction accuracy (Table 2). Statistical comparison using Student’s t-test revealed no significant differences between the predicted and ground truth measurements for all evaluated traits, indicating high consistency and precision of the YOLOv8 derived measurements. The MAE, as defined by Willmott and Matsuura (2005), serves as a critical metric for quantifying discrepancies between predicted and reference values⁴⁰. Notably, the MAE values for fruit width, fruit length, peel thickness, and number of fruit segments were low, with YOLOv8 achieving an MAE of approximately 0.3–0.9 pixels, highlighting its ability to accurately capture fine-scale structural details. Similarly, for fruit area, peel area, segment area, and central core area, which are the key quantitative traits in plant breeding research, the model achieved MAE values of 800, 4449, 4472, and 2345 pixel², respectively, demonstrating strong agreement with the ground truth annotations.

Table 2.

Sample comparative analysis of mandarin sliced parameters from YOLOv8 predictions and ground truth.

Traits	Ground truth	Predicted	T-test	MAE
Fruit width	1564.61 ± 10.08	1586.85 ± 10.10	0.99	0.30
Fruit length	1638.35 ± 10.55	1638.62 ± 10.10	0.99	0.30
Fruit area	2,064,682.75 ± 26,200.95	2,065,408.65 ± 26,286.53	0.98	800
Peel area	305,808.82 ± 6844.38	304,080.25 ± 6836.48	0.86	4449
Peel thickness	62.16 ± 1.15	61.77 ± 1.15	0.81	0.90
Segments area	1,640,001.95 ± 20,450.51	1,640,726.37 ± 20,488.51	0.98	4472
Number of fruit segments	11.64 ± 0.11	11.40 ± 0.12	0.16	0.30
Central core area	118,871.98 ± 3477.86	120,602.03 ± 3503.15	0.73	2345

When considered alongside the nonsignificant t-test results, these findings confirm the superior accuracy and robustness of YOLOv8 in predicting morphological characteristics of sliced mandarin fruits. The model effectively captures intricate anatomical features, thereby enhancing the reliability of quantitative fruit morphology analyses. Overall, the YOLOv8 architecture shows strong potential for high-throughput phenotyping, cultivar evaluation, and scalable quantitative fruit morphology research through precise and automated image-based measurements.

Discussion

Quantitative fruit characteristics are fundamental descriptors of citrus fruit development and quality^5,39. Morphological diversity analysis is particularly important because many key agronomic traits are controlled by multiple genes, and their additive effects can be evaluated through phenotypic expression^5,41. Furthermore, assessing genetic diversity using morphological traits, such as fruit width, fruit length, fruit area, number of fruit segments, fruit segment area, and central core area, is advantageous due to the availability of standardized descriptors and the ease of visual inspection^7,12. Large-scale evaluation of citrus fruit anatomical traits can eliminate knowledge gaps through systematic phenotypic characterization, as limited genotype-level knowledge of useful and diverse morphological traits across citrus species persists, thereby enhancing citrus improvement and the selection of desirable morphological characteristics in breeding programs^14,30,32.

In this study, we applied YOLOv8 to detect and segment fruit segments and the central core in transversely cut mandarin fruit images. The results demonstrate strong potential for large-scale application, enabling precise and accurate extraction of biologically relevant anatomical characteristics. Moreover, unlike many previous studies that emphasize detection or classification performance alone, our approach integrates accurate instance segmentation with quantitative trait validation, ensuring that the extracted measurements retain biological and phenotypic relevance. The high mean IoU values (> 95% for fruit segments and > 80% for the central core) indicate accurate delineation of anatomical boundaries, which is essential for reliable downstream trait extraction. Similarly, the high precision, recall, mAP, and F1-score values confirm that the model achieves a robust balance between minimizing false positives and false negatives. These results compare favorably with previous deep learning-based fruit segmentation studies, which often report lower segmentation accuracy for complex internal structures (Muresan and Oltean 2018; Lin et al. 2021).

The native YOLOv8 framework was selected because it provides an effective balance between accuracy, speed, and robustness, making it well suited for high-throughput agricultural phenotyping without unnecessary model complexity. Additionally, our controlled in-laboratory imaging setup ensured clear class separability from the background, allowing the architecture to perform optimally. We further strengthened the study by including comparative experiments with the newer YOLOv11 model to provide a more comprehensive evaluation (Table 3).

Table 3.

Performance comparison between YOLOv8 and YOLOv11 on mandarin slice segmentation.

Metric	YOLOv8	YOLOv11
Box Precision (P)	0.997	0.996
Box Recall (R)	1.0	1.0
Box mAP50	0.995	0.995
Box mAP50–95	0.922	0.925
Mask Precision (P)	0.997	0.996
Mask Recall (R)	1.0	1.0
Mask mAP50	0.965	0.962
Inference time
(ms/image)	84.4 ms	8.1 ms
Postprocess time
(ms/image)	2.8 ms	6.1 ms

A key strength of the present study lies in the comprehensive measurement of biologically meaningful internal quantitative anatomical traits of mandarin fruit, which reflect the coordinated processes of cell division, expansion, and tissue differentiation during fruit growth. The nonsignificant differences between YOLOv8-predicted measurements and ground truth data, together with low mean absolute error (MAE) values, demonstrate that the model not only performs well in segmentation metrics but also preserves measurement fidelity. Even for area-based traits, which are more sensitive to boundary inaccuracies, the achieved MAE values indicate strong agreement with manual annotations. The particularly low MAE observed for fruit width, fruit length, and segment number highlights the model’s ability to capture fine-scale geometric details, which is crucial for cultivar discrimination and genetic studies.

Moreover, in commercial practice, fruit shape and size underpin automated fruit grading, and market classification systems, where consistency and objectivity of measurements are critical^24,29,42. In mandarin, segment number and segment area are closely associated with ovary locule number and vascular development, whereas the central core area reflects internal structural organization and can influence both seed number and variation in consumer preference among different species or cultivars^2,4. These traits can be distinguished through their morphological characteristics and representative quantitative parameters. Consequently, these parameters are widely used in plant breeding and cultivar evaluation to discriminate genotypes, assess genetic diversity, and select superior cultivars with desirable size, shape, and internal structure¹¹.

Although the morphological parameters in this study were reported in pixel units, they can be readily converted into real-world measurements. In the imaging setup, a label tag with known physical dimensions (in millimeters) is visible in the original scanned images and can serve as a reference for spatial calibration. By measuring the size of this reference object in pixels within the image and comparing it with its known physical size, a pixel-to-millimeter conversion factor can be calculated. This calibration allows all extracted traits, such as segment width, core diameter, and radial distances, to be converted from pixel units to millimeters if required. Therefore, while pixel-based measurements were used for computational consistency in the present analysis, the proposed framework can easily provide biologically meaningful dimensions suitable for fruit phenotyping and quantitative morphological studies.

Because this study focuses on achieving high-precision results for applications in agricultural and quantitative fruit morphology research, several limitations should be acknowledged. The dataset size, although well annotated and statistically balanced, remains relatively small and was collected under controlled imaging conditions using a flatbed scanner and a uniform background. Variations in lighting intensity, shadow effects, and complex or cluttered backgrounds were not included in the present dataset, which may limit the robustness of the model under real-world acquisition scenarios. In addition, the current framework has not yet been validated using portable capture devices such as mobile phones, which may introduce differences in resolution, color response, perspective distortion, and illumination variability. Although the instance segmentation model achieved high accuracy, occasional false detections or segmentation inaccuracies may still occur, particularly for smaller or less visually distinct structures such as the central core, which could subsequently influence the accuracy of the derived morphological measurements. Furthermore, the proposed analytical pipeline involves several sequential processing steps, including instance segmentation, contour extraction, ellipse fitting, radial distance computation, and peak detection, which increases methodological complexity and may introduce cumulative errors if earlier stages are affected by segmentation noise. The lower IoU observed for the central core compared with fruit segments further highlights the challenge of accurately segmenting small anatomical structures.

Future studies should therefore incorporate larger, multi-cultivar datasets collected under diverse lighting conditions and heterogeneous backgrounds, as well as cross-device validation using portable imaging systems. The development of a standardized yet portable image acquisition setup potentially including a mobile capture platform with color calibration references and illumination control would facilitate deployment in field or industrial environments. Cross-institutional validation and domain adaptation strategies should also be explored to further assess and enhance model generalizability. Despite these limitations, the results clearly demonstrate that YOLOv8 is a robust and effective tool for the automated detection, segmentation, and quantitative analysis of sliced mandarin fruit characteristics, supporting its application in high-throughput phenotyping, breeding programs, and precision fruit quality assessment.

Conclusions

In this paper, the YOLOv8 model is applied as an automated method for segmentation and measurement of mandarin fruit morphology. The results demonstrate that the YOLOv8 instance segmentation framework enables accurate detection, segmentation, and quantitative analysis of anatomical structures in sliced mandarin fruit images. The well-annotated and statistically validated dataset supported robust model training, capturing diverse morphological patterns of fruit segments and the central core. YOLOv8 achieved high segmentation and detection performance, with strong IoU, precision, recall, and F1-score values, indicating reliable delineation of anatomical regions and stable model convergence. Furthermore, a Python-based algorithm was developed to extract several anatomical parameters, including fruit area, fruit width, fruit length, number of fruit segments, segment area, and central core area. The YOLOv8 derived measurements showed close agreement with ground truth values and low prediction error, highlighting the model’s ability to accurately capture fruit morphological features. Overall, the results confirm that YOLOv8 is a robust and reliable tool for automated analysis of mandarin fruit morphology, with strong potential for high-throughput phenotyping and image-based applications in agricultural research. The implications of this study are significant for fruit breeding, cultivar evaluation, and the validation of quantitative fruit characteristics. Collectively, this work represents a meaningful advancement in understanding plant structural complexity and contributes to both theoretical knowledge and practical applications in agriculture and plant breeding.

Abbreviations

IoU

Intersection over union

MAE

Mean absolute error

RGB

Red, green and blue

YOLOv8

You only look once version 8

Author contributions

JA supervised the study, provided guidance, edited the final version, and acquired funding. TTT designed the model and wrote and edited the manuscript. ATL curated the data and wrote the original draft of the manuscript. All authors contributed to the final version of the manuscript.

Funding

This work was supported by the the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00348897).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval and consent to participate

Not applicable.

Consent for publication

All authors have reviewed the final manuscript and consent to its publication. No individual person’s data are included that would require additional consent.