Enhanced object detection of Enterobius vermicularis eggs using cumulative transfer learning algorithm

Related works

Table 1 provides a comprehensive comparison of related works with other studies, specifically analyzing their conclusions, identified research gaps, and methodological limitations. This comparison underscores the novelty of our research, which introduces an innovative approach for the detection of E. vermicularis eggs utilizing the scotch tape technique. By addressing the limitations of previous studies, our proposed method enhances diagnostic accuracy and efficiency, contributing to advancements in parasitological detection techniques.

Data collection

This study used E. vermicularis egg images from a publicly available dataset accessible through Figshare (https://doi.org/10.6084/m9.figshare.26266028.v2) (Chaibutr et al., 2024). The E. vermicularis eggs were collected using the established scotch tape method. A piece of transparent adhesive tape was gently pressed against the perianal region of patients, typically first thing in the morning before bathing or defecation, to collect nocturnal egg deposits. The tape was then affixed onto a glass slide and examined microscopically to confirm the presence of characteristic E. vermicularis eggs. Confirmed eggs were photographed using a digital microscope camera for subsequent image processing and analysis. Each image was manually inspected and annotated by experienced laboratory personnel. The annotated images were separated into two distinct classes: (1) class 0 for negative samples or artifact structures with 1,000 images, and (2) class 1 for positive samples of 1,000 images exhibiting the characteristic morphology of E. vermicularis eggs (Figs. 1A, 1B). This manual curation process ensured that only high-quality, correctly labeled samples were included in the training and testing sets. The dataset comprised 2,000 high-resolution (2,448 × 1,920 pixels) Tagged Image File Format images with balanced distribution between E. vermicularis eggs and artifacts. For the training and validation, we utilized the Google Colab platform with robust computational resources: an NVIDIA A100 GPU, Intel Xeon CPU operating at 2.3 GHz, 83.5 GB of RAM, and 235.7 GB of disk space.

The study protocol was approved by the Ethics Committee in Human Research, Walailak University (protocol code WUEC-24-387-01, approved 8 November 2024) and followed the Declaration of Helsinki. The dataset was distributed under the Creative Commons Attribution (CC BY) license and contained only microscopic images of artifacts and E. vermicularis eggs, with no personal identifiable information or sensitive data.

Dataset preparation

Generalizability and robustness of the trained model, the original dataset underwent several advanced image augmentation procedures. Gaussian blur was introduced to replicate variations in focus commonly observed in real-world microscopic imaging, thereby enabling the model to handle slightly out-of-focus samples. Mean filtering was employed to smooth images and mitigate high-frequency noise, while Gaussian noise addition mimicked the random intensity fluctuations frequently encountered in microscopic observations. Finally, kernel sharpening was applied to emphasize edges and subtle morphological details of the eggs, ensuring that essential diagnostic features remained discernible (Figs. 1C, 1D). By integrating these augmentation techniques, the dataset’s diversity was substantially increased, resulting in improved model performance and enhanced resilience to image quality variations commonly encountered in clinical applications. This approach enables the model to maintain robust detection capabilities when processing unseen data with varying imaging conditions, such as differences in microscope settings, illumination levels, and sample preparation protocols. The enhanced dataset diversity ensures that the trained model can generalize effectively across different laboratory environments and imaging equipment configurations (Chaibutr et al., 2024; Lundin et al., 2024; Thanchomnang et al., 2024).

Model construction and learning

In this study, a five-fold cross-validation protocol was employed to rigorously evaluate model performance, ensuring that each subset of the dataset served as both training and validation data in the separate runs for a more reliable measure of generalizability. The network was trained for a maximum of 200 epochs unless the mean-squared-error (MSE) loss fell below 0.001, which acted as an early stopping criterion (Tuite et al., 2011). The model architecture comprised the three sequential blocks of convolutional layers of decreasing dimension in terms of filter count. Each block contained three consecutive Conv2D layers with ReLU activation initially 64 filters in the first block, then 32 filters in the second, and finally 8 filters in the third block, followed by a MaxPooling2D operation to reduce spatial dimensionality. The feature maps were then flattened and passed through four fully connected (Dense) layers, sized 256, 64, 16, and two neurons, respectively. Each layer employed the sigmoid activation function to progressively distill the essential features and ultimately output the classification results. The network was compiled using a stochastic gradient descent (SGD) as the optimizer and the MSE loss function, with accuracy tracked as the primary performance metric. This configuration aimed to balance computational efficiency with sufficient modeling capacity, leveraging progressive feature extraction in the convolutional layers and feature refinement in the dense layers for robust egg detection.

Conventional training (Fig. 2A) in a supervised learning context typically began with importing the dataset from a designated source, such as an image repository or a curated collection of labeled samples. Once imported, data augmentation is performed to artificially expand and diversify the dataset by applying transformations such as rotation, flipping, or slight color modifications, while preserving the essential features necessary for accurate classification or detection. This augmented dataset was then fed into a selected model architecture, where training was conducted to optimize the model’s parameters through iterative updates driven by a chosen loss function and optimization algorithm. Finally, the validation phase was performed using a portion of the data that was held out, enabling the assessment of the model’s performance and generalization capabilities. This conventional pipeline seeks to balance the model’s ability to learn robust representations with the need to avoid overfitting, ultimately aiming to achieve high accuracy on unseen data (Miseta, Fodor & Vathy-Fogarassy, 2024).

Cumulative transfer learning (Fig. 2B) refines the conventional supervised learning pipeline encompassing data collection, labeling, augmentation, and model training by integrating a sequential, knowledge-preserving methodology (Yang et al., 2019). The process begins with the assembly of an annotated dataset, which is then augmented to enhance the breadth of feature patterns and mitigates overfitting. Rather than instantiating model weights from scratch, training commences by importing those from a previously optimized model, thereby exploiting established feature representations (Liu et al., 2020). Crucially, the incoming augmented dataset was divided into smaller subsets, each introduced to the model incrementally, allowing the network to acquire new insights while preserving earlier learned parameters. Between each incremental phase, we performance the evaluations via metrics such as accuracy, precision, or recall to ensure the newly assimilated data enhances rather than overwrites existing knowledge (i.e., mitigating the issue of “catastrophic forgetting”). Over successive phases, this adaptive inheritance of learned representations yielded robust performance, improved generalization, and accelerated convergence relative to conventional single-stage training. Such a tiered approach is especially advantageous for domains afflicted by limited or imbalanced datasets, as it maximizes the utility of previously gleaned insights, fostering more nuanced feature extraction and heightened classification accuracy in dynamic or evolving tasks (Zorić et al., 2024). By incorporating knowledge transfer at multiple stages, this method allows the model to capture nuanced morphological features of E. vermicularis eggs, including elliptical structure, shell thickness, and internal granularity. These subtle characteristics are often overlooked by traditional methods (Xu et al., 2016).

Cumulative transfer learning, step-by-step:

Assume we have three datasets

$$DatasetA:D_A={\left\{\left(x_A^i,y_A^i\right)\right\}}_{i=1}^{N_A}$$

(1) $$DatasetB:D_B={\left\{\left(x_B^i,y_B^i\right)\right\}}_{i=1}^{N_B}$$

$$DatasetC:D_C={\left\{\left(x_C^i,y_C^i\right)\right\}}_{i=1}^{N_C}$$where $x^i$ represents an input sample, $y^i$ is the corresponding label, N is the number of samples in each dataset.

Step 1: Pre-training on Dataset A

The first model is trained to minimize a loss function ${\zeta }_A$

(2) $${\theta }_A^{\ast }=argmin{\mathrm{E}}_{\left(x,y\right)\sim P_A}\left[{\zeta }_A\left(f\left(x;\theta \right),y\right)\right]$$

Using gradient descent algorithm (Eq. (3)):

(3) $${\theta }_A^{\ast }={\theta }_A-\eta {\displaystyle \frac{\mathrm{\partial }{\zeta }_A}{\mathrm{\partial }\theta }}$$where ${\theta }_A^{\ast }$ represent the optimal parameters for dataset A, η is learning rate, ${\displaystyle \frac{\mathrm{\partial }{\zeta }_A}{\mathrm{\partial }\theta }}$ is the gradient loss function.

Step 2: Fine-tuning on Dataset B

We initialize the model for Dataset B with ${\theta }_A^{\ast }$, instead of random initialization

(4) $${\theta }_B^{\ast }=argmin{\mathrm{E}}_{\left(x,y\right)\sim P_B}\left[{\zeta }_B\left(f\left(x;\theta \right),y\right)\right].$$

Using gradient descent algorithm on 2^nd transfer learning iteration (Eq. (5)):

(5) $${\theta }_B^{\ast }={\theta }_A^{\ast }-\eta {\displaystyle \frac{\mathrm{\partial }{\mathrm{\zeta }}_B}{\mathrm{\partial }\theta }.}$$

Step 3: Fine-tuning on Dataset C

We initialize the model for Dataset B with ${\theta }_A^{\ast }$, instead of random initialization

(6) $${\theta }_C^{\ast }=argmin{\mathrm{E}}_{\left(x,y\right)\sim P_C}\left[{\zeta }_C\left(f\left(x;\theta \right),y\right)\right].$$

Using gradient descent algorithm on 3^rd transfer learning iteration (Eq. (7)):

(7) $${\theta }_C^{\ast }={\theta }_B^{\ast }-\eta {\displaystyle \frac{\mathrm{\partial }{\zeta }_C}{\mathrm{\partial }\theta }.}$$

The algorithms, codes, and README file are available at the following GitHub repository: https://github.com/Pongphan/ctransferlearning_ev (accessed on 22 March 2025).

Evaluation metrics

A comprehensive set of evaluation metrics was employed to assess model performance. Accuracy measured the proportion of correctly classified images relative to the total number of samples. Precision quantified the proportion of positive identifications that were truly positive, indicating the reliability of positive predictions. Recall measured the model’s ability to identify all relevant instances, ensuring minimal misses true positives (TPs). The F1-score, as the harmonic mean of precision and recall, provided a balanced performance measure particularly useful for class imbalance scenarios. Sensitivity and specificity were included for their clinical significance. Sensitivity measured the ability to detect TPs, while specificity evaluated the capacity to correctly identify true negatives. These metrics provided a multi-dimensional understanding of model performance, with particular emphasis on minimizing false negatives (FNs) due to their critical importance in clinical diagnostics.

Comparison to baseline models

To assess the efficacy and robustness of our proposed cumulative transfer learning framework, we conducted a comprehensive comparative analysis against several widely recognized deep neural network architectures that commonly employ in medical image classifications. Specifically, we selected conventional CNN (Chaibutr et al., 2024), ResNet50, InceptionV3, DenseNet121, and Xception as benchmark models due to their established track records in handling complex visual patterns, their proven capability to achieve high accuracy on large-scale image recognition tasks, and their broad adoption within the medical imaging community (Gunturu et al., 2024). Conventional CNN, the architecture begins with a sequence of convolutional layers, each utilizing 64 filters with a 3 × 3 kernel size. After every two convolutional layers, max pooling is applied to downsample the spatial dimensions of the feature maps, thereby reducing computational complexity and emphasizing the most critical features. To further prevent overfitting, dropout is introduced following the max-pooling layers, randomly omitting units during training. The model then progresses from convolutional operations to fully connected layers, comprising three dense layers with 128, 64, and eight units, respectively. Each dense layer is also followed by batch normalization and dropout to maintain regularization. The final dense layer contains two units, making it appropriate for binary classification tasks. This architecture, characterized by its depth and the integration of multiple regularization techniques, is designed to effectively manage complex data while minimizing the risk of overfitting (Chaibutr et al., 2024). ResNet50, with its residual connections, mitigates the issues related to the vanishing gradients and has been shown to excel in extracting detailed hierarchical features (Rahmati, Shirani & Keshavarz-Motamed, 2024). InceptionV3’s inception modules promote an efficient parallel feature extraction across multiple scales, while DenseNet121’s densely connected layers facilitate feature reuse, ultimately reducing model parameters without compromising performance (Hu et al., 2021). Similarly, Xception’s depthwise separable convolutions are enable a more efficient factorization of convolutional operations, potentially improving both accuracy and computational efficiency (Peng & Wang, 2021). By comparing our approach against these influential and architecturally distinct networks, we aimed to establish a robust benchmark and discerned the relative advantages were conferred by our cumulative transfer learning strategy in enhancing the sensitivity, specificity, and overall diagnostic reliability of E. vermicularis egg detection. To compare the performance of object detection models, IoU is commonly used as a standard metric. IoU quantifies the spatial overlap between a predicted bounding box and the corresponding ground truth annotation, calculated as the area of intersection divided by the area of union between the two boxes. This provides an objective measure of localization accuracy.

For comparative analysis, we also implemented the You Only Look Once version 8 (YOLOv8) object detection framework as a baseline. YOLOv8 is a state-of-the-art, single-stage object detection model renowned for its speed and accuracy in diverse computer vision tasks (Ko & Lee, 2025). The model was trained on the same annotated dataset comprising images of E. vermicularis eggs and artifacts, following identical preprocessing and augmentation protocols as used for the cumulative transfer learning approach. Training parameters including input image size, learning rate, and batch size were optimized to ensure fair comparison. Detection performance was quantitatively evaluated using IoU metrics, with predicted bounding boxes compared against expert-annotated ground truth on the test set. This benchmarking enabled direct assessment of the relative strengths and limitations of the YOLOv8 model in the context of automated parasitological diagnostics.

Dataset

All images used for training were standardized by resizing them to 128 × 128 pixels to ensure uniformity across the dataset. Initially, 1,000 images were collected for each of the two classes namely, artifacts (class 0) and E. vermicularis eggs (class 1) resulting in a balanced starting dataset of 2,000 images. To enhance the diversity and robustness of the training data, a multi-step augmentation procedure was implemented, beginning with random rotations followed by sequential application of a mean filter, Gaussian blur, Gaussian noise, and sharpening. Each of these transformations was systematically applied to increase the dataset by a factor of four for each original image, ultimately yielding 4,000 augmented images per class and contributing to improved model generalization. For evaluation, an independent testing set comprising 100 images of artifacts (class 0) and 100 images of E. vermicularis eggs (class 1) was curated under the same resolution constraints. This clearly separated test set enabled an unbiased assessment of the trained model’s performance on previously unseen data.

Conventional learning

After completing the five-fold cross-validation procedure, in which the dataset is partitioned into five equally sized subsets such that each subset serves as the validation set exactly once, Fig. 3A showed line plots of classification accuracy for each fold. Error bars were included to illustrate variability across partitions, thereby demonstrating the model’s performance consistency under different data splits. In the same Fig. 3B, the receiver operating characteristic (ROC) curve illustrated the trade-off between sensitivity (TP rate) and 1—specificity (FP rate), with a diagonal reference line indicating a random classification. The area under the ROC curve (ROC-AUC) quantified the model’s discriminative power, where values approaching 1.0 signify high diagnostic efficacy. Taken together, the fold-specific accuracy trends and ROC analysis provided a comprehensive assessment of the model’s robustness and predictive capability under varying training validation configurations.

Cumulative transfer learning

In our cumulative transfer learning paradigm, the training process was subdivided into five consecutive rounds, each determined by the augmented data generated at each stage. Within each round, we applied a five-fold cross-validation, partitioning the dataset into five unique subsets to evaluate model performance and enhance generalization ability. Figures 4 and 5 presented the outcomes of this iterative five-fold cross-validation from round 0 to round 5, encompassing the entire progression of the cumulative transfer learning approach. Each point reflects the model’s average performance across the five validation folds, providing robust estimates of classification accuracy, loss, and other relevant metrics. The figures are organized into three key sections: (1) cross-validation loss for each fold, (2) accuracy per fold, and (3) the ROC curves. This framework illustrates the model’s learning trajectory in each round and demonstrates improvements in both classification performance and generalizability. The cross-validation loss plots highlight how the model’s error evolved across progressive rounds. Lower loss values in later rounds indicated enhanced ability to detect E. vermicularis eggs, reflecting successful knowledge transfer in cumulative learning. The middle section presents the accuracy achieved in each of the five folds, showing how classification performance improves from round 0 to round 5. Rising accuracy from one round to the next underscores the positive impact of incremental transfer learning on the feature extraction and classification tasks. The rightmost section in Fig. 5 shows ROC curves, providing an in-depth view of the TP rate against the FP rate for each fold. By comparing curve shapes and their corresponding area under the curve (AUC) values, researchers can evaluate both the sensitivity and specificity gains throughout the cumulative transfer learning process. Collectively, these panels offer comprehensive performance evaluation. Figure 5 underscores the robustness of the proposed cumulative transfer learning approach. Moreover, the five-fold cross-validation design to ensure that the reported performance is not overly dependent on any particular data partition, thereby reinforcing the reliability of these findings for clinical applications.

As illustrated in Table 2, a detailed comparison between conventional CNN and cumulative transfer learning revealed a modest, yet appreciable improvement in model performance, as evidenced by a slight increase in overall accuracy when employing the cumulative transfer strategy. Beyond the accuracy gain, a significant reduction in training time was observed: the cumulative transfer learning model required only 0:51:44 h for completion, compared to the 2:07:32 h needed by the conventional approach. This acceleration largely contributed to more efficient parameter initialization and the progressive reuse of knowledge across training phases, thereby minimizing redundancy in feature extraction.

In addition to these time savings, the cumulative transfer learning also demonstrated a marked decrease in resource consumption, where the peak memory usage dropped from 56.7 to 19.4 GB. Such a substantial reduction was not only lowers computational costs but can also expanded the feasibility of deploying advanced DL models in resource-constrained settings. Collectively, these findings underscored the potential value of cumulative transfer learning in both enhancing diagnostic accuracy and optimizing computational efficiency for E. vermicularis egg detection.

In this study, we conducted a comparative evaluation of four DL architectures ResNet50, InceptionV3, DenseNet121, and Xception as summarized in Table 3. The goal was to assess each model’s performance in classifying E. vermicularis eggs acquired from scotch tape samples. Of the tested architectures, two models emerged with superior metrics: our proposed the cumulative transfer learning CNN and the InceptionV3 network. Notably, both systems achieved a perfect classification performance, demonstrated by an accuracy, precision, recall, and F1-score of 1.0. These results underscored the effectiveness of cumulative transfer learning, which enabled iterative refinement across multiple training phases, thereby leveraging knowledge from prior training steps to enhance final predictive capabilities. Furthermore, the notable success of InceptionV3 recognized for its multi-branch convolution modules and dimensional reduction techniques reinforced that the advanced DL methodologies can outperform traditional baselines in detecting parasitic eggs. Achieving perfect evaluation metrics in this diagnostic setting highlights the promise of these approaches for reliable, high-throughput laboratory implementation.

Object detection

A comprehensive comparative analysis of object detection outputs generated by the high-precision ML models, compared against expert medical annotations on microscopic images, is presented in Fig. 6. By juxtaposing these two sets of annotations, one can directly evaluated the model’s detection accuracy, including its ability to localize and identify target structures, as well as to identify any discrepancies that may inform further refinement of the detection algorithm. In Fig. 6A, red bounding box represented the expert-verified ground truth, enabling clear visualization of clinically relevant target. The automatically generated bounding boxes are highlighted in green from the proposed model (Fig. 6B) and YOLOv8 model (Fig. 6C).

An in-depth overview of the IoU thresholds obtained by the proposed model was displayed in Table 4. Comparative analysis of the proposed cumulative transfer learning CNN and YOLOv8 reveals distinct performance characteristics. The proposed cumulative transfer learning approach achieved perfect sensitivity (TP = 6, FN = 0) but exhibited a high FP rate (FP = 97). Its average precision (AP) @IoU0.5 (0.530) substantially outperformed YOLOv8 (0.245), indicating superior detection precision at the standard IoU threshold. However, its mean average precision (mAP) @IoU0.5:0.05:0.95 (0.027) was considerably lower than YOLOv8 (0.057), suggesting limited generalization across stricter IoU thresholds. In contrast, YOLOv8 demonstrated higher detection capability (TP = 23) with fewer FPs (FP = 71) but suffered from missed detections (FN = 9) and lower precision at IoU0.5. The superior mAP of YOLOv8 reflects greater robustness across varying overlap criteria, likely attributable to its anchor-free architecture and optimized localization-classification balance. These results indicate that the proposed cumulative transfer learning model prioritizes sensitivity and single-threshold precision, while YOLOv8 demonstrates superior overall localization stability. The findings suggest that model selection should be guided by application-specific requirements regarding tolerance for FNs vs localization accuracy demands.