Detection of Schistosoma Eggs Using an AI-Based Deep Learning Model on Urinary Bladder Histopathology Images

To assess the accuracy and clinical applicability of YOLO-based segmentation models for detecting Schistosoma eggs in urinary bladder histopathology slide images, focusing on both bounding box and mask segmentation tasks.

A descriptive, cross-sectional study. Place and Duration of the Study: Artificial Intelligence Technology Centre, National Centre for Physics, Islamabad, Pakistan, from September to November 2024.

A high-quality dataset was compiled using histopathological slides obtained from real patient samples available on the open-source platform PathPresenter. All images were meticulously annotated by expert histopathologists. The dataset included 681 images containing 2,751 schistosomes, divided into 476 training images (1,932 schistosomes), 136 validation images (539 schistosomes), and 69 testing images (280 schistosomes). Data pre-processing techniques were applied to optimise the quality of training and evaluation datasets. Multiple YOLO-based segmentation models, such as YOLOv5, YOLOv8, YOLOv9, and YOLOv11 variants (n/s/m/l/x/c/e), were trained and evaluated for both bounding box and mask detection. Model performance was evaluated using precision, recall, F1 score, mean Average Precision at 50% Intersection over Union (IoU; mAP50), and mAP across 50 to 95% IoU (mAP50-95) for both bounding box and mask segmentation tasks.

Among the models, YOLOv8l demonstrated the highest diagnostic accuracy, achieving an F1 score of 95.09 and a mAP50 of 96.8 for bounding box detection. For mask detection, it attained an F1 score of 94.19 and an mAP50 of 96.2. YOLOv5m and YOLOv5x also performed well, balancing accuracy with computational efficiency. Smaller models exhibited limitations in sensitivity and precision.

YOLO-based segmentation models exhibit strong potential for automated detection of schistosomiasis in urinary bladder histopathology images. However, future large-scale validation studies on bigger datasets are required for further confirmation.

Deep learning, Mask segmentation, Medical image analysis, Schistosomiasis, YOLO segmentation.