Abstract

Traditional diagnostic methods in medical parasitology rely heavily on manual microscopic examination, which is labor-intensive and prone to human error and subjectivity. This study introduced a novel approach for automating the detection of Enterobius vermicularis (pinworm) eggs using cumulative transfer learning algorithms. The proposed framework effectively captures subtle egg morphology by employing a sequential knowledge transfer paradigm, thereby enhancing diagnostic accuracy, efficiency, and reproducibility, even when data are limited. This study used E. vermicularis egg images from a publicly available dataset. The training image dataset comprised 1,000 images of artifacts and 1,000 images of pinworm eggs. Comparisons were made against established deep learning (DL) models, including conventional convolutional neural network (CNN), ResNet50, DenseNet121, Xception, and InceptionV3. Results demonstrated that the cumulative transfer learning strategy consistently outperformed both the conventional CNN method and DL baselines in terms of classification accuracy, F1-score, and computational efficiency, while also reducing computational overhead. Performance comparison with a conventional CNN model demonstrates that the proposed cumulative transfer learning CNN reduces training time from 2 h to 50 min. Moreover, it achieves optimal performance, with accuracy, precision, recall, and F1-score all reaching 1.0. The model’s detection accuracy was quantitatively assessed by comparing predicted bounding boxes to expert annotations across 103 microscopic images. The proposed cumulative transfer learning CNN achieved higher average precision (AP) @ intersection over union (IoU) 0.5 (0.530) and perfect sensitivity (1.00), but exhibited 97 false positives and lower mean average precision (mAP) @IoU0.5:0.05:0.95 (0.027). In contrast, the You Only Look Once version 8 (YOLOv8) model demonstrated lower sensitivity (0.72) but superior multi-threshold performance (mAP@IoU0.5:0.05:0.95 = 0.057). These results highlight a trade-off between detection sensitivity and generalization performance across varying IoU thresholds. These findings affirm the viability of cumulative transfer learning as a scalable, accurate, and efficient approach for automated parasitological diagnostics, particularly in resource-limited settings.