Fig 1.
Example images of the prediction results of the trained model on unseen data.
(a) – (f) Example images from the “white pearl” demonstrate the capacity of the algorithm to detect and classify positive and negative results of individual wells in 8-tube LAMP strips. The model demonstrates robust detection performance across different illumination conditions and color variations. Green and red bounding boxes indicate negative and positive LAMP reactions, respectively, with confidence scores above each detection. The model maintains high detection accuracy (confidence scores ranging from 0.73 to 0.91) despite challenging variations in: (1) background illumination (purple to brown tones), (2) fluorescence intensity (bright to dim), (3) sample opacity, and (4) image capture angles.
Fig 2.
Training and validation metrics for the YOLOv8 model on LAMP image detection.
The plots show the progression of various loss functions (box, class, and dfl losses) for both training and validation sets for 80 epochs, with all metrics displayed on the same figure. Performance metrics based on the validation set, including precision, recall, mAP@50, and mAP@50-95 are also displayed, demonstrating the model’s improving accuracy in detecting and classifying LAMP assay results. The smooth convergence of loss functions and the upward trends in accuracy metrics indicate successful training and good generalization of the model for LAMP image analysis.
Table 1.
Performance comparison of the YOLOv8 model between the validation and test sets.
Table 2.
Comparative performance analysis of YOLOv8, YOLOv7, and Faster R-CNN models on the test set for LAMP image classification.
Fig 3.
Confusion matrix demonstrating high-accuracy classification of LAMP images.
Fig 4.
Performance Metrics (a) Precision-Confidence, (b) Recall-Confidence, (c) Precision-Recall, (d) F1-Confidence.
Fig 5.
The input image shows a LAMP test result with fluorescent reactions. The YOLO architecture consists of three main components: (a) the backbone for hierarchical feature extraction at different scales, capturing both fine and coarse details, (b) the neck (green circles with bidirectional dotted arrows) for multi-scale feature aggregation and enhancement. The bidirectional arrows indicate feature fusion and information flow between different scales allowing the model to better understand multi-scale features. (c) the head (blue diamonds) for final predictions: object classification and bounding box regression. Having three heads suggest that predictions are made at three different scales which is typical in YOLO architectures for detecting objects of various sizes. The three rows represent different feature map scales for detecting various object sizes: top(small), middle (medium), and bottom (large).
Fig 6.
The architecture of YOLOv8 consists of a backbone, neck, and head, adopted from [31].
Fig 7.
White Pearl LAMP device.
Fig 8.
Training dataset examples for YOLOv8 algorithm to detect and classify LAMP assay results.
Images depict 8-tube LAMP strips with positive results fluorescing bright green and negative results as a dull orange. (a) and (c) are images captured on a commercial transilluminator. (b) and (d) are images captured using the white pearl transilluminator.