BlotDx: A deep learning tool for Western blot-based diagnostics

BlotDx: A deep learning tool for Western blot-based diagnostics | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article BlotDx: A deep learning tool for Western blot-based diagnostics Lucas Liu, Saraswathi Sathees, Aniela Sobel, Gregory Pepper, Alexander L. Greninger, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6605358/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Western blot (WB) is the gold standard for herpes simplex virus (HSV-1/HSV-2) serology but requires manual interpretation by trained laboratory personnel, which is time-consuming and variable. Automating this process could improve efficiency and diagnostic consistency. This study presents BlotDx, a deep learning tool for automated interpretation of HSV WB assays. Methods BlotDx uses a two-stage approach: (1) instance segmentation or object detection to identify blot strips from input images, and (2) classification models to determine HSV-1 and HSV-2 serostatus. The dataset used contained 34 images, each with up to 24 pairs of blot strips. Of these, 21 images were used for training, 4 for validation, and 9 for testing (n = 314, 57, and 124 blot pairs respectively). Data augmentation, ensemble modeling, and transfer learning were used to enhance model performance. Findings Compared to the gold standard of expert human review by three independent MLSs, BlotDx demonstrated high diagnostic accuracy (99% for HSV-1, 98% for HSV-2) across validation and test datasets. Data augmentation and ensemble methods significantly improved robustness and reliability. The transfer learning approach allowed the models to leverage pre-trained weights, further enhancing performance. Interpretation This study highlights the utility of AI in automating the analysis of Western blots for HSV diagnostics, with potential applications in other diseases. The proposed two-stage approach, combined with modern deep learning techniques, achieved scalable, accurate, and efficient diagnostic results. These findings underscore the potential of deep learning to transform traditional diagnostic workflows by reducing costs and increasing efficiency. deep learning Western blots herpes simplex virus diagnostics clinical testing serology Figures Figure 1 Figure 2 Figure 3 INTRODUCTION Deep learning and artificial intelligence (AI), particularly through convolutional neural networks (CNNs) and vision transformers, have revolutionized image-based analysis, offering solutions to automate labor-intensive tasks and standardize results [ 1 , 2 ]. In medical diagnostics, these advancements hold promise for improving accuracy and efficiency. CNNs are now often used in radiology and pathology [ 3 , 4 ], in combination with transfer learning [ 5 ], to help with disease detection, classification, and segmentation. However lab-based diagnostics that involve image interpretation (e.g. Western blot-based assays) have seen relatively little application of deep learning to date. Western blot assays, widely employed in clinical diagnostic and research applications, are analytical techniques used to detect specific proteins by analyzing visual band patterns. The process involves separating proteins based on their molecular weight through gel electrophoresis, transferring them onto a membrane, and probing the membrane with antibodies. The presence or absence of antibodies to specific proteins is then visualized as distinct bands, verifying the presence of disease-specific markers or confirming the immune response in a patient sample [ 6 , 7 ]. For example, the Western blot assays used for diagnosing herpes simplex virus (HSV) infection detect the presence of specific antibodies against HSV proteins in a patient's serum [ 5 ]. These type-specific assays can also determine the subtype of the virus and consist of two separate blots: one with HSV-1 proteins and the second with HSV-2 proteins, resulting in the detection of HSV-1 and/or HSV-2 antibodies in the patient’s serum. The ability to distinguish between HSV-1 and HSV-2 antibodies is achieved by the unique protein profiles associated with each virus and accurate interpretation of these bands is critical for diagnostic accuracy. HSV is a globally prevalent, sexually transmitted virus that causes painful lesions and chronic, lifelong infection. HSV diagnosis can be challenging because lesions may be absent at the time of clinical evaluation, and the frequency of recurrence can be variable across individuals. Serological assays can be used to determine whether the person was previously exposed to HSV even if there are no active lesions present and the results can be important for medical decision-making, communication to partners, family planning, and public health efforts [ 8 ]. Early enzyme-linked immunoassays (ELISAs, [ 9 ]) did not distinguish between HSV-1 and HSV-2 and were shown to be inaccurate, particularly in detecting HSV-2 antibodies in HSV-1 seropositive persons [ 10 , 11 ]. Type-specific assays based on glycoprotein-G [ 12 , 13 ] became commercially available in the late 1980s and were shown to be more accurate [ 14 ]. However, tests based on antibodies to a single protein are more vulnerable to suboptimal sensitivity and specificity [ 15 ], as well as to antibody titers and timing, since titers tend to rise 2–3 months after infection in most cases, and sometimes up to 6 months after infection [ 16 ]. As a result, HSV Western blot assays continue to be the gold standard due to their ability to more definitively distinguish between HSV-1 and HSV-2, and also to identify seroconversion earlier [ 16 ]. When used for clinical testing, Western blot-based assays can be resource-intensive, as they require trained technicians to interpret results manually. The variability in human interpretation can lead to inconsistencies, particularly in ambiguous or borderline cases. Deep learning has demonstrated potential in automating such analyses in other medical imaging applications [ 17 – 19 ]. The primary goal is to design a system that achieves the accuracy of trained technicians while minimizing the need for manual interpretation. This study introduces convolutional neural network (CNN)-based models for automating the segmentation and classification of Western blot images using HSV serological testing as a use case. The study focuses on automating the analysis of Western blot images by addressing two key objectives. First, the process of cropping blot strips from the original images must be automated to ensure accurate segmentation. Second, the cropped strips must be robustly classified into diagnostic categories. As this was a small pilot study, a relatively simple set of images was used where a diagnosis was obtained after the first round of testing (positive or negative for HSV-1 and HSV-2 separately). Future enhancements will include indeterminate results and edge cases so that BlotDx is capable of handling the full range of variability seen in Western blot images. Common challenges in applying CNNs for Western blot-based diagnostics include variability in image quality, differences in band patterns, and limited availability of annotated datasets. This study aims to overcome these obstacles by integrating robust preprocessing, data augmentation, and model optimization techniques. METHODS Dataset: HSV Western blot images and clinical test results The dataset for this study comprised 34 multi-subject images, each containing up to 24 pairs of strips, with each pair corresponding to HSV-1 and HSV-2 serological assay result for a single patient serum sample. The images were generated using a clinically validated HSV Western blot assay (UW-WBA, https://testguide.labmed.uw.edu/view/HSWB , last accessed 23-Mar-25). The UW-WBA is currently the gold-standard method for HSV serology due to its high sensitivity and ability to distinguish between HSV-1 and HSV-2 infection and was performed using methods described previously [ 11 , 20 ]. Briefly, HSV-1 and HSV-2 proteins from infected cell lysates were separated by electrophoresis and transferred on to nitrocellulose paper. The resulting strips of paper (blots) were incubated with patient serum and antibodies that bind to the viral proteins were detected by an enzyme-mediated color change. The resulting banding patterns on the blots are read by three independent MLSs, and the consensus used to call the final result as positive, negative, or indeterminate for HSV-1 and HSV-2 antibodies. Both type-specific and type-common antibodies to HSV-1 and HSV-2 can be detected in Western blot assays. Type-specific antibodies are directed exclusively against either HSV-1 or HSV-2 antigens, whereas type-common antibodies recognize proteins shared between both virus types [ 21 ]. For example, a type-common antibody produced in response to an HSV-1 infection may bind to proteins on the HSV-2 blot strip, potentially leading to a false impression of dual infection. As a result, approximately 20% of patient serum samples may require additional steps to clarify the results and accurately determine the infecting virus type. To resolve these ambiguous cases, a pre-adsorption technique is employed. This involves incubating the patient's serum with HSV-1 protein lysate that has been bound to Sepharose beads. During incubation, antibodies specific to HSV-1 as well as type-common antibodies that cross-react with HSV-1 antigens will bind to the lysate and be effectively removed from the sample. The adsorbed serum is then tested using the HSV Western blot assay. Theoretically, only HSV-2–specific antibodies will remain in the sample, allowing for a clearer interpretation of HSV-2 infection status. Similarly, a parallel adsorption procedure is performed using HSV-2 protein lysate coupled to Sepharose beads. The final result is then called by three MLSs reading the post-adsorption blots. In this pilot study, only blots that had a positive or negative result for HSV-1 and HSV-2 were included, and indeterminate blots and blots needing repeat testing or adsorption were excluded for the classification step. Adsorption blots were only used for the blot identification step (Stage 1 described in the next section). The images were divided into 21 for training (n = 314 blot pairs), 4 for validation (n = 57 blot pairs), and 9 for testing (n = 124 blot pairs). In addition, for instance segmentation, one additional image containing 7 groups of 6 strips (from adsorptions) was included in the training set to improve training data diversity (Supplementary Figure S1). Ground truth bounding boxes and masks for training object detection and instance segmentation models were drawn using the VGG Image Annotator [ 22 ]. Ground truth labels, i.e. test results for each strip, were obtained from clinical results in the form of a table containing manual interpretation by 3 technicians, and a final result following supervisory review. The final result was used to create the ground truth labels. BlotDx: a computer vision and deep learning tool for automating the interpretation of HSV Western blots A two-stage framework (Fig. 1 ) was adopted to achieve the diagnostic goal of predicting HSV serostatus from input images containing multiple pairs of strips. The first stage crops blot strips from the original images via instance segmentation or object detection methods, and the second stage classifies the cropped strips into diagnostic categories (HSV-1-HSV-2-, HSV-1 + HSV-2-, HSV-1-HSV-2+, HSV-1 + HSV-2+,),. The two-stage framework was chosen over a one-stage framework (where segmentation/detection and classification are carried out in the same model using the original images as input) because features important for cropping strips may be different from features important for classifying strips. Stage 1: Detecting or segmenting strips using deep learning models To detect pairs of strips corresponding to individual subjects, improved Faster R-CNN models with a ResNet-50-FPN backbone were used [ 23 – 25 ] with the number of classes set to 2. To segment individual strips, improved Mask R-CNN models with a ResNet-50-FPN backbone were used with the number of classes set to 2. For both cases, the loss function was adjusted to halve the weight of the classification loss in the total loss computation. Model training was performed by stochastic gradient descent with learning rate 0.001, momentum 0.9, weight decay 0.0005, for 1000 epochs with a batch size of 4. Standard data augmentations including random horizontal and vertical flip, random scaling from 50–150%, and random rotation of up to 360 degrees were applied. Transfer learning was implemented by starting model training from weights trained on the Common Objects in Context (COCO) dataset. Image processing for classification Additional image processing is needed on the outputs of Stage 1 in order to proceed to classification. The pre-processing required depends on whether instance segmentation or object detection was used for stage 1. If instance segmentation was performed, it outputs masks for the strips which require processing with two main steps: 1) A quadrilateral was fit to each mask. The top edge of the quadrilateral was determined as a horizontal line passing through the 20th percentile of the y-coordinates of the uppermost points in the pair of masks corresponding to the individual containing the input mask. The bottom edge was set as a horizontal line positioned 420 pixels below the top edge. The target rectangle height was chosen such that no bands in the strips were left out. The left edge was estimated by fitting a robust linear regression model to the leftmost points of the mask, while the right edge was estimated similarly using the rightmost points of the mask. 2) The fitted quadrilateral was used to crop out the blot image and transform it into a 23 × 420-pixel rectangle using an affine shearing transformation, which preserves the y-coordinate of every point in the trapezoid. The target rectangle width was chosen as the 95th percentile of the strip widths. Once cropped, the two rectangular strips from the same individual were combined either side-by-side to create a 3 x 46 x 420 tensor or stacked top-and-bottom to form a 6 x 23 × 420 tensor. If object detection was performed in stage 1, it outputs a rectangular bounding box of varying size. A simple pre-processing step is performed to crop the image using the bounding box and rescale the cropped image containing the double strips to be 46 pixels wide and 420 pixels tall. Stage 2: Classification using deep learning models For classification, a ResNet-50 model was used to predict HSV-1 and HSV-2 status (positive or negative) separately. Due to cross-reactivity between antibodies against HSV-1 and HSV-2, all classifiers were designed to take both strips as input. Three classifiers were designed (SEG1, SEG2, and DET) varying in how the two strips were provided to the classifier (Fig. 1 ). Because the strips may be slanted to different degrees in the original image, SEG1 and SEG2 were designed to segment out the two strips separately, transform each into a rectangular shape, and combine them. The relative relationship between the bands on the two strips was maintained during the transformation. The difference between SEG1 and SEG2 lies in how the transformed strips were arranged—side-by-side in SEG1 and stacked vertically in SEG2. DET was designed as an alternative to SEG1 and SEG2, opting not to "correct" the slanting of the strips in the original images. Instead, it employed a bounding box to crop both strips and rescaled the cropped region to a fixed dimension. Model training was performed by stochastic gradient descent with learning rate 0.001, momentum 0.9, and weight decay 0.0005, for 100 epochs with a batch size of 24. Data augmentations including random scaling from 90–110% and random rotation of up to 2 degrees were applied. Transfer learning was implemented by starting model training from weights trained on the IMAGENET1K dataset. Classification model saliency maps To gain insights into how the classification models operate, we generate saliency maps [ 26 ] using multiple gradient-based methods including class activation maps (CAM) [ 27 ], Grad-CAM [ 28 ], and Grad-CAM++ [ 29 ]. These widely used model interpretation methods generate saliency maps by computing weighted combinations of feature maps from convolutional layers of the classification model, which are then up-sampled to the input image resolution using bilinear interpolation. The weights for combining the feature maps reflect each map's contribution to the predicted class and are based on the gradients of the score for a class with respect to the map. For CAM, the gradients are taken with respect to the feature vector after the global average pooling layer. For Grad-CAM, gradients are computed with respect to each pixel in the feature vectors in the final convolutional layer and averaged across the spatial dimensions. Grad-CAM + + generalizes Grad-CAM by replacing simple gradient averaging with a weighted average of the positive gradients. These gradient weights are formally derived by expressing the classification score as a linear combination of the feature vectors after global average pooling. Statistical Analyses All statistical analyses were performed using R 4.2.2 (R Project for Statistical Computing). Code and Data Availability The python and R scripts developed for this project are available in the GitHub repository at https://github.com/youyifong/BlotDx . A snapshot of the repository, taken at the time of this writing, is archived in Zenodo at https://doi.org/10.5281/zenodo.15284041 . The multi-subject images that are the input to Stage 1 contain subject-identifying information and cannot be shared publicly. The single-subject images that are the input to Stage 2 are free of subject-identifying information and available on Zenodo at https://doi.org/10.5281/zenodo.15067008 . Classification model checkpoints are available from the same Zenodo repository. Role of the funding source No external funding was received for this study. RESULTS Stage 1: Object detection and instance segmentation methods accurately identify blot strips from multi-subject images The first stage of our diagnostic tool BlotDx involves identification of blot strips using either object detection or instance segmentation. The two methods were compared in their effectiveness to correctly identify all pairs of blots given an input image with multiple pairs, with each pair containing an HSV-1 (left) strip and HSV-2 (right) strip (Figure 2). The object detection method returns bounding boxes surrounding each pair of strips for a single subject, while the instance segmentation method returns an instance mask, which is converted to a bounding box for individual strips for each subject. For object detection-based inference, at the minimum confidence score of 0.5 for predicted bounding boxes, all pairs of strips (n=87, 100%) were identified by bounding boxes in the validation set, but many small spurious boxes were also predicted. These were present mainly near the lower boundary of the image possibly due to shadows/artifacts present in the input image (example in Figure S2). Raising the threshold to 0.8 did not remove all the unwanted boxes. Keeping the box confidence score at 0.5 and adding a minimum size for a bounding box to be retained at 1000 pixels removed all unwanted boxes in both the validation set (n = 57 blot pairs) and the test set (n = 124 blot pairs). For instance segmentation-based inference, using a minimum confidence score of 0.5 for predicted bounding boxes failed to identify all strips in the validation set. Lowering the confidence threshold to 0.1 and introducing a minimum bounding box size of 5000 pixels ensured the detection of all strips without generating unwanted masks. However, multiple masks were occasionally predicted for the same strip. This was addressed by applying an Intersection over Union (IoU) threshold of 0.7 for non-maximum suppression within the Region Proposal Network. Applying these criteria resulted in the detection of all individual strips with no spurious masks in both the validation and test set. The quality of the inferred masks and bounding boxes was determined by comparing them against ground-truth masks and bounding boxes (GT) for the prediction of serostatus in the validation set (n = 57) using the three classifiers developed (SEG1, SEG2, DET, see Methods and Figure 1). Each classifier was executed three times using different seeds for the random number generator and a majority vote call was determined across the three results. All three classifiers achieved >94% accuracy, and performance was equivalent to or better than classification using GT labels (P-value 0.5, 0.5, 1 for SEG1, SEG2, DET respectively), indicating the adequacy of the inferred masks and bounding boxes for the purposes of classification (Table 1). Table 1. Classification accuracy of the three classifiers SEG1, SEG2, and DET using inferred vs. ground truth (GT) masks or bounding boxes based on number of correctly predicted results in the validation set (n = 57 blot pairs) over three replicate runs and an overall number calculated by the majority call. Stage 2: Accurate identification of HSV-1 and HSV-2 status using individual and ensemble classifiers The performance of the three classifiers using inferred masks and bounding boxes was evaluated in the validation (n = 57) test sets (n = 124), with results from 3 replicate runs combined through majority vote. The three classifiers (SEG1, SEG2, and DET) were evaluated individually and an ensemble approach was also tested that combined the three results through majority voting. For HSV-1, SEG1 and SEG2 achieved 98.4% and 95.2% accuracy in the test set respectively, while DET achieved perfect classification (Table 2). An ensemble of the three methods achieved accuracies of 98.2% and 99.2% in the validation and test sets respectively. Pooling the results over the validation set and the test sets, the ensemble method achieved a 98.9% accuracy (n = 181). The two misclassified samples were HSV-1 positive misclassified as negative (Figure S3). We repeated the training of classifiers without transfer learning, i.e. starting from random weights instead of ImageNet-trained weights. For the ensemble method, forgoing transfer learning led to 5 additional HSV-1 positive samples misclassified as negative with no gains. The McNemar's test comparing results with and without transfer learning yielded a P-value of 0.0625, suggesting that models benefited from pre-trained weights. For HSV-2, the ensemble method achieved 97.8% (177/181) accuracy overall (97.2% for SEG1, 97.8% for SEG2, 97.2% for DET, Table 2). The four misclassified samples included two that were HSV-2 positive misclassified as negative and two that were misclassified in the opposite direction. Starting model training from random weights led to the level of accuracy for the ensemble method. Table 2. Classification accuracy of the three classifiers SEG1, SEG2, and DET on the test set (n = 124) and a combined set of validation and test images (n = 181) with and without transfer learning (TL). With TL No TL Test (n = 124) Combined (n = 181) Combined (n = 181) P-value* SEG1 HSV-1 122 178 175 0.25 HSV-2 119 176 177 1.00 SEG2 HSV-1 118 177 168 0.01 HSV-2 120 177 173 0.22 DET HSV-1 124 178 174 0.22 HSV-2 119 176 177 1.00 Ensemble HSV-1 123 99.2% 179 98.9% 174 96.1% 0.06 HSV-2 120 96.8% 177 97.8% 177 97.8% 1.00 *McNemar’s test Overall, when implemented for both subtypes with transfer learning and using an ensemble method, BlotDx produced 98.9% test accuracy of, 97.6% sensitivity, and 100% specificity for HSV-1, and 97.8% test accuracy, 95.5% sensitivity, and 98.5% specificity for HSV-2 (Table 3). Table 3. Confusion matrices for ensemble results with transfer learning HSV-1 HSV-2 True\Predicted Negative Positive Negative Positive Negative 97 0 134 2 Positive 2 82 2 43 To obtain a more robust estimate of classification performance, we conducted a 5-fold cross-validation experiment. All blot pairs were randomly partitioned into five folds. Models were trained on four folds and evaluated on the remaining fold. The ensemble approach achieved classification accuracies of 98.8% for HSV-1 and 98.6% for HSV-2 (Table S1). The slight performance improvement may be attributed to the higher training-to-test data ratio in the 5-fold cross-validation setup. Paired HSV-1 and HSV-2 blots provide improved classification accuracy compared to single blots In the models above, due to expected cross-reactivity between HSV-1 and HSV-2 antigens (typically referred to as type-common antibodies), all classifiers were designed to take HSV-1 and HSV-2 strips for each subject together as input. To assess the impact of cross-reactivity on classification accuracy, these paired-strip classifiers were compared against an alternative single-strip design. In the single strip design, the HSV-1 result was generated by a classifier that only received the HSV-1 strip as input, and likewise for the HSV-2 classifier. The single-strip classifiers achieved a lower overall classification accuracy (93.4% for HSV-1 and 97.2% for HSV-2, Table 4), suggesting that information from both blots is important for classification. Table 4. Classification accuracy of a single-strip classifier using SEG1 on the validation set (n = 57) and test set images (n = 124) with and without transfer learning (TL). With TL No TL Validation Test Validation Test HSV-1 52 117 50 111 HSV-2 57 119 57 119 Classification model saliency maps Saliency maps were used to visualize the relative importance of different spatial regions to the classification result. Four different algorithms were compared (CAM, GradCAM, and GradCAM++, see Methods). Comparing the HSV-1 and HSV-2 saliency maps (Figure 3), we see that the areas that contribute to HSV-2 positive identification are limited to the HSV-2 strip, while the areas that contribute to HSV-1 positive identification appear to spread between both the HSV-1 strip and the HSV-2 strip. This supports our previous result that showed that both blot strips are required for accurate classification. The saliency maps from the HSV-1 models also show greater image-to-image variability than those from the HSV-2 model. The four different algorithms had good visual correspondence with each other, but there were some differences in the importance regions for SEG1 and DET classifiers, particularly for HSV-1 models (Figure 3A). For the prediction of HSV-1+/HSV-2- status, high significance regions for HSV-1 models (Figure 3A(i)) included bands corresponding to HSV-2 glycoproteins gB/gG, gC, gE, and major capsid protein VP5 [30,31] for both SEG1 and DET classifiers. The assembly protein ICP35 region was highlighted in both HSV-1 and HSV-2 blots for the SEG1 classifier, but only on the HSV-2 blot for the DET classifier. For the prediction of HSV-1+/HSV-2+ status by the HSV-1 models (Figure 3A(iii)), the SEG1 classifier focused primarily on the gB/gG region of the HSV-1 blot, while the DET classifier focused on the gD region of the HSV-2 blot. The HSV-2 models showed fewer differences between SEG1 and DET classifiers and in all cases the highlighted regions of high importance for both single and double-positive cases (HSV-1-/HSV-2+ and HSV-1+/HSV-2+ in Figure 3B) we on the HSV-2 blot alone. Of note, both SEG1 and DET classifiers for HSV-2 models strongly highlighted the gB/gG region, which is consistent with the way the blots are interpreted in the lab where a strong gG band on the HSV-2 blot is a key requirement for an HSV-2 positive result. DISCUSSION This study presents an AI-based tool (BlotDx) for the analysis of diagnostic Western blots. While AI-based methods have been applied to Western blots in other settings like plagiarism detection [ 32 ], BlotDx is, to our knowledge, the first such application of CNNs for serodiagnostics in a clinical laboratory setting. We present a two-stage deep learning solution that performs identification of individual or paired blot strips from multiple-sample input images and then performs classification to return a result for HSV-1 and HSV-2 serostatus. The proposed two-stage solution identified all strips from multi-subject images and achieved high diagnostic accuracy for HSV-1 and HSV-2. Given that the current laboratory workflow involves interpretation by multiple technicians followed by final supervisory review, BlotDx offers opportunities for cost savings and efficiency by potentially simplifying the process to require fewer reviewers. Although currently designed for HSV serological assays, the same approach could be easily adapted for other applications that employ Western blots, e.g. for Lyme disease [ 33 , 34 ], human T-lymphotrophic virus [ 35 ], bovine spongiform encephalopathy [ 36 ], and tapeworm infections like Echinococcus [ 37 ]. The findings also highlight the value of transfer learning and ensemble methods, and of combining object detection and instance segmentation models in a two-stage process for improved model performance and stability. While modern computer vision models can simultaneously identify and classify objects, two-stage approaches that separate identification from classification have demonstrated improved performance in various practical applications, such as cancer diagnosis [ 38 ], skin lesion analysis [ 39 ], and splenomegaly classification [ 40 ]. As a step towards explainable AI, we used saliency maps to identify regions of importance within the input image, highlighting which pixels or regions contributed the most towards a classification result. The finding that the HSV-1 model’s saliency maps encompass signal from both HSV-1 and HSV-2 blots may be related to our finding that single-strip classifiers underperform compared to classifiers that use both strips as input. One possibility is that the model may have implicitly learned that information from the HSV-2 blot is predictive of the HSV-1 label—either due to biological correlation or experimental artifact—and thus relies on both regions for its decision-making. The second explanation is that this phenomenon reflects an instance of aliasing as described by Li et al.[ 41 ], where spatially distinct but statistically correlated features introduce misleading attributions in saliency maps. This could occur if regions within the HSV-1 and HSV-2 blots exhibit high inter-feature correlation (perhaps in a subset of positive training samples), causing the gradients to erroneously attribute relevance across both regions. The third possibility is that the saliency maps, derived from gradients of the final feature extraction layer, represent higher-level semantics and fail to localize the spatial source of discriminative signal accurately. This explanation, however, appears less compelling given that the saliency maps generated from the HSV-2 model are confined to the HSV-2 blot, suggesting that spatial localization is at least partially preserved in the saliency pipeline. Results from saliency maps are also consistent with the way the blots are read in the lab. MLSs interpreting the blots use information from both strips, and particular attention is paid to glycoprotein G which, when present as a strong band on the HSV-2 blot indicates a high probability of HSV-2 infection. HSV-1 and HSV-2 are closely related herpesviruses with > 70% sequence identity, reflecting their shared evolutionary origin [ 42 – 45 ]. This relatedness gives rise to both type-common and type-specific antigens. Western blot assays detect a spectrum of antibody responses, producing a mixture of shared and unique banding patterns. Key type-specific antigens—such as glycoprotein G-1 (gG-1) for HSV-1 and glycoprotein G-2 (gG-2) for HSV-2—enable differentiation between the two subtypes when interpreted correctly. However, reliance on gG alone has shown to be less sensitive to detecting seropositivity when compared to using the full spectrum of bands in the UW-WBA [ 20 ]. While this work demonstrates the potential utility of AI in lab-based diagnostics, our study had a number of limitations. As this was a pilot study, a relatively small dataset was used and strips with indeterminate results that required re-testing or adsorption were excluded. Data augmentation was used to mitigate the small size of the dataset to some extent but a larger dataset with greater complexity, and classification with more result classes would improve the usability in a practical setting. As the images were obtained by photographing archived blot strips, some fading is expected over time, and this may have contributed to some of the misclassifications. The use of freshly run blots, and photographing strips on a dark background to enhance contrast may lead to further improvement in classification accuracy. The black-box nature of CNNs raises concerns about interpretability, and future work should aim to integrate tools that provide insights into model decisions to enhance trust among clinicians. There is also potential for incorporating the image capture and analysis into an automated benchtop solution that accelerates the sample-to-answer process. Overall, this work paves the way for AI-powered solutions that bridge the gap between traditional diagnostics and modern computational methods. References LeCun Y., Bengio Y., Hinton G. Deep learning. Nature. 2015;521: 436–444. PMID 26017442 Dosovitskiy A., Beyer L., Kolesnikov A., Weissenborn D., Zhai X., Unterthiner T., Dehghani M., Minderer M., Heigold G., Gelly S., Uszkoreit J., Houlsby N. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv [cs.CV]. 2020. Available: http://arxiv.org/abs/2010.11929 Litjens G., Kooi T., Bejnordi B.E., Setio A.A.A., Ciompi F., Ghafoorian M., van der Laak J.A.W.M., van Ginneken B., Sánchez C.I. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42: 60–88. PMID 28778026 Esteva A., Chou K., Yeung S., Naik N., Madani A., Mottaghi A., Liu Y., Topol E., Dean J., Socher R. Deep learning-enabled medical computer vision. NPJ Digit Med. 2021;4: 5. PMID 33420381PMCID PMC7794558 Kim H.E., Cosa-Linan A., Santhanam N., Jannesari M., Maros M.E., Ganslandt T. Transfer learning for medical image classification: a literature review. BMC Med Imaging. 2022;22: 69. PMID 35418051PMCID PMC9007400 Towbin H., Staehelin T., Gordon J. Electrophoretic transfer of proteins from polyacrylamide gels to nitrocellulose sheets: procedure and some applications. Proc Natl Acad Sci U S A. 1979;76: 4350–4354. PMID 388439PMCID PMC411572 Begum H., Murugesan P., Tangutur A.D. Western blotting: a powerful staple in scientific and biomedical research. Biotechniques. 2022;73: 58–69. PMID 35775367 Workowski K.A., Bachmann L.H., Chan P.A., Johnston C.M., Muzny C.A., Park I., Reno H., Zenilman J.M., Bolan G.A. Sexually transmitted infections treatment guidelines, 2021. MMWR Recomm Rep. 2021;70: 1–187. PMID 34292926PMCID PMC8344968 Kimmel N., Friedman M.G., Sarov I. Enzyme-linked immunosorbent assay (ELISA) for detection of herpes simplex virus-specific IgM antibodies. J Virol Methods. 1982;4: 219–227. PMID 6286702 Ashley R., Cent A., Maggs V., Nahmias A., Corey L. Inability of enzyme immunoassays to discriminate between infections with herpes simplex virus types 1 and 2. Ann Intern Med. 1991;115: 520–526. PMID 1652909 Wald A., Ashley-Morrow R. Serological testing for herpes simplex virus (HSV)-1 and HSV-2 infection. Clin Infect Dis. 2002;35: S173-82. PMID 12353203 Katz D., Hilliard J.K., Mirkovic R.R., Word R.A. ELISA for detection of IgG and IgM antibodies to HSV-1 and HSV-2 in human sera. J Virol Methods. 1986;14: 43–55. PMID 3021796 Lee F.K., Coleman R.M., Pereira L., Bailey P.D., Tatsuno M., Nahmias A.J. Detection of herpes simplex virus type 2-specific antibody with glycoprotein G. J Clin Microbiol. 1985;22: 641–644. PMID 3001136PMCID PMC268483 Ashley R.L. Sorting out the new HSV type specific antibody tests. Sex Transm Infect. 2001;77: 232–237. PMID 11463920PMCID PMC1744343 Crawford K.H.D., Selke S., Pepper G., Goecker E., Sobel A., Wald A., Johnston C., Greninger A.L. Performance characteristics of highly automated HSV-1 and HSV-2 IgG testing. J Clin Microbiol. 2024;62: e0026324. PMID 38687020PMCID PMC11237592 Ashley R.L., Wald A. Genital herpes: review of the epidemic and potential use of type-specific serology. Clin Microbiol Rev. 1999;12: 1–8. PMID 9880471PMCID PMC88903 Pinto-Coelho L. How artificial intelligence is shaping medical imaging technology: A survey of innovations and applications. Bioengineering (Basel). 2023;10. PMID 38136026PMCID PMC10740686doi: 10.3390/bioengineering10121435 Liu X., Faes L., Kale A.U., Wagner S.K., Fu D.J., Bruynseels A., Mahendiran T., Moraes G., Shamdas M., Kern C., Ledsam J.R., Schmid M.K., Balaskas K., Topol E.J., Bachmann L.M., Keane P.A., Denniston A.K. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health. 2019;1: e271–e297. PMID 33323251 Varoquaux G., Cheplygina V. Machine learning for medical imaging: methodological failures and recommendations for the future. NPJ Digit Med. 2022;5: 48. PMID 35413988PMCID PMC9005663 Ashley R.L., Militoni J., Lee F., Nahmias A., Corey L. Comparison of Western blot (immunoblot) and glycoprotein G-specific immunodot enzyme assay for detecting antibodies to herpes simplex virus types 1 and 2 in human sera. J Clin Microbiol. 1988;26: 662–667. PMID 2835389PMCID PMC266403 Bernstein D.I., Bryson Y.J., Lovett M.A. Antibody response to type-common and type-unique epitopes of herpes simplex virus polypeptides. J Med Virol. 1985;15: 251–263. PMID 2580054 Dutta A., Zisserman A. The VIA Annotation Software for Images, Audio and Video. Proceedings of the 27th ACM International Conference on Multimedia. New York, NY, USA: ACM; 2019. doi: 10.1145/3343031.3350535 Li Y., Xie S., Chen X., Dollar P., He K., Girshick R. Benchmarking detection transfer learning with vision transformers. arXiv [cs.CV]. 2021. Available: http://arxiv.org/abs/2111.11429 Ren S., He K., Girshick R., Sun J. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans Pattern Anal Mach Intell. 2017;39: 1137–1149. PMID 27295650 He K., Gkioxari G., Dollar P., Girshick R. Mask R-CNN. 2017 IEEE International Conference on Computer Vision (ICCV). IEEE; 2017. pp. 2980–2988. Szczepankiewicz K., Popowicz A., Charkiewicz K., Nałęcz-Charkiewicz K., Szczepankiewicz M., Lasota S., Zawistowski P., Radlak K. Ground truth based comparison of saliency maps algorithms. Sci Rep. 2023;13: 16887. PMID 37803108PMCID PMC10558518 Zhou B., Khosla A., Lapedriza A., Oliva A., Torralba A. Learning deep features for discriminative localization. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2016. doi: 10.1109/cvpr.2016.319 Selvaraju R.R., Cogswell M., Das A., Vedantam R., Parikh D., Batra D. Grad-CAM: Visual explanations from deep networks via gradient-based localization. Int J Comput Vis. 2020;128: 336–359. Chattopadhay A., Sarkar A., Howlader P., Balasubramanian V.N. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE; 2018. doi: 10.1109/wacv.2018.00097 McKendall R.R., Pettit M., Woo W. The polyspecific immunoglobulin response to HSV-1 viral proteins: determination of immunogenic proteins and relative antibody titres to individual polypeptides by immunoblotting. J Med Microbiol. 1988;25: 49–57. PMID 2826789 Wales S.Q., Smith C.C., Wachsman M., Calton G., Aurelian L. Performance and use of a ribonucleotide reductase herpes simplex virus type-specific serological assay. Clin Diagn Lab Immunol. 2004;11: 42–49. PMID 14715543PMCID PMC321330 Gosselin R.-D. AI detectors are poor western blot classifiers: a study of accuracy and predictive values. PeerJ. 2025;13: e18988. PMID 39989748PMCID PMC11847483 Dressler F., Whalen J.A., Reinhardt B.N., Steere A.C. Western blotting in the serodiagnosis of Lyme disease. J Infect Dis. 1993;167: 392–400. PMID 8380611 John T.M., Taege A.J. Appropriate laboratory testing in Lyme disease. Cleve Clin J Med. 2019;86: 751–759. PMID 31710588 Ribeiro L.C.P., Gonçalves C.C.A., Slater C.M.S.A., Carvalho S.M.F. de, Puccioni-Sohler M. Standardisation of Western blotting to detect HTLV-1 antibodies synthesised in the central nervous system of HAM/TSP patients. Mem Inst Oswaldo Cruz. 2013;108: 730–734. PMID 24037195PMCID PMC3970696 Okada H., Murayama Y., Shimozaki N., Yoshioka M., Masujin K., Imamura M., Iwamaru Y., Matsuura Y., Miyazawa K., Fukuda S., Yokoyama T., Mohri S. Prion in saliva of bovine spongiform encephalopathy–infected cattle. Emerg Infect Dis. 2012;18: 2091–2092. Deininger S., Wellinghausen N. Evaluation of a new combined Western and line blot assay (EUROLINE-WB) for diagnosis and species identification of Echinococcus infection in humans. GMS Infect Dis. 2019;7: Doc01. PMID 30993059PMCID PMC6446614 Madhu G., Meher Bonasi A., Kautish S., Almazyad A.S., Mohamed A.W., Werner F., Hosseinzadeh M., Shokouhifar M. UCapsNet: A two-stage deep learning model using U-Net and Capsule Network for breast cancer segmentation and classification in ultrasound imaging. Cancers (Basel). 2024;16. PMID 39594732PMCID PMC11592068doi: 10.3390/cancers16223777 Innani S., Dutande P., Baheti B., Baid U., Talbar S. Deep learning based novel cascaded approach for skin lesion analysis. InInternational Conference on Computer Vision and Image Processing. Cham; Nature Switzerland: Springer; 2022. pp. 615–626. Meddeb A., Kossen T., Bressem K.K., Molinski N., Hamm B., Nagel S.N. Two-stage deep learning model for automated segmentation and classification of splenomegaly. Cancers (Basel). 2022;14: 5476. PMID 36428569PMCID PMC9688308 Li K., Wu Z., Peng K.-C., Ernst J., Fu Y. Tell me where to look: Guided attention inference network. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE; 2018. doi: 10.1109/cvpr.2018.00960 Aguilar J.S., Devi-Rao G.V., Rice M.K., Sunabe J., Ghazal P., Wagner E.K. Quantitative comparison of the HSV-1 and HSV-2 transcriptomes using DNA microarray analysis. Virology. 2006;348: 233–241. PMID 16448680 Baines J.D., Pellett P.E. Genetic comparison of human alphaherpesvirus genomes. In: Arvin A, Campadelli-Fiume G, Mocarski E, Moore PS, Roizman B, Whitley R, et al., editors. Human Herpesviruses. Cambridge: Cambridge University Press; 2007. pp. 61–69. Schmid D.S. Mixing it up: New insights into interspecies recombination between herpes simplex virus type 1 and 2. The journal of infectious diseases. Oxford University Press (OUP); 2020. pp. 1208–1209. PMID 31107962PMCID PMC9137430 Casto A.M., Roychoudhury P., Xie H., Selke S., Perchetti G.A., Wofford H., Huang M.-L., Verjans G.M.G.M., Gottlieb G.S., Wald A., Jerome K.R., Koelle D.M., Johnston C., Greninger A.L. Large, Stable, Contemporary Interspecies Recombination Events in Circulating Human Herpes Simplex Viruses. J Infect Dis. 2020;221: 1271–1279. PMID 31016321PMCID PMC7325804 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6605358","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":452830304,"identity":"6a2d5cd6-6e65-44d9-9fff-c864bf57d1be","order_by":0,"name":"Lucas Liu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Lucas","middleName":"","lastName":"Liu","suffix":""},{"id":452830305,"identity":"1e7a0fbc-0930-4d01-88f8-566b4ebffeb4","order_by":1,"name":"Saraswathi Sathees","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Saraswathi","middleName":"","lastName":"Sathees","suffix":""},{"id":452830306,"identity":"c6c14240-b024-4acd-ad14-7e9ce6eb050c","order_by":2,"name":"Aniela Sobel","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Aniela","middleName":"","lastName":"Sobel","suffix":""},{"id":452830307,"identity":"e66ca81c-d218-4584-b929-12b600a0d3f5","order_by":3,"name":"Gregory Pepper","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Gregory","middleName":"","lastName":"Pepper","suffix":""},{"id":452830308,"identity":"d1d5094a-d927-44d3-b825-b097007c01ec","order_by":4,"name":"Alexander L. Greninger","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alexander","middleName":"L.","lastName":"Greninger","suffix":""},{"id":452830309,"identity":"5a963ba9-775a-406b-98ca-9967845ae4ab","order_by":5,"name":"Keith R. Jerome","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Keith","middleName":"R.","lastName":"Jerome","suffix":""},{"id":452830310,"identity":"de35c982-00db-4262-89da-aa3d5194e5b5","order_by":6,"name":"Youyi Fong","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Youyi","middleName":"","lastName":"Fong","suffix":""},{"id":452830311,"identity":"c4f8e8e6-05a6-460f-8169-547c001f23a7","order_by":7,"name":"Pavitra Roychoudhury","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAqklEQVRIiWNgGAWjYLACxgYGBn4wi40ULZINJGsxOECsFnPpw8ckfu6wSdx8I8eA4UPZYcJaLPvS0iR7z6QlbgNqYZxxjggtBmd4zKQZ2w4DteRuYOZtI0XL5hlALX9J0rJBAqiFkRgtlj1syZa9bWnGM868/3Cw51w6YS3mPMwHb/xss5Htb09LfPCjzJoIhzEwsEjAOAcIq4doYf5AlMpRMApGwSgYuQAADYQ79i7Cj0UAAAAASUVORK5CYII=","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Pavitra","middleName":"","lastName":"Roychoudhury","suffix":""}],"badges":[],"createdAt":"2025-05-06 17:14:34","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6605358/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6605358/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82359881,"identity":"de9ec093-681f-4b6e-947d-b5607b221943","added_by":"auto","created_at":"2025-05-09 11:34:12","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":553142,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic of BlotDx, a computer vision and deep learning-based tool to assist in clinical diagnosis of HSV-1 and HSV-2 seropositivity.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6605358/v1/a3d06ccbbbb37e574aee575a.png"},{"id":82358891,"identity":"1573ce28-116e-473e-91a8-f4467859bdd5","added_by":"auto","created_at":"2025-05-09 11:26:12","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":969467,"visible":true,"origin":"","legend":"\u003cp\u003eSample images with predicted object detection (left) and instance segmentation (right) masks. Identifying information has been redacted with light blue bands below the blot strips.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6605358/v1/38e7c5871ca1321a9bbf401d.png"},{"id":82359882,"identity":"2a47c3ea-ad53-437e-87ce-845e362045a6","added_by":"auto","created_at":"2025-05-09 11:34:12","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3423463,"visible":true,"origin":"","legend":"\u003cp\u003eClassification model saliency maps for three blot pairs that were correctly classified by both DET and SEG1 classifiers for HSV-1 (A) and HSV-2 (B). Test images: (i) HSV-1 positive and HSV-2 negative, (ii) HSV-1 negative and HSV-2 positive, (iii) HSV-1 positive and HSV-2 positive. For each panel, the input image is shown on the left followed (L-R) by saliency maps from CAM, GradCAM, and GradCAM++ respectively.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6605358/v1/e758fdfd4220a59b6cde8b07.png"},{"id":82361384,"identity":"fac2c4d2-8d54-4651-8b1e-a07efdbfa0fa","added_by":"auto","created_at":"2025-05-09 11:50:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5628368,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6605358/v1/72a15ff9-cbec-49e9-8a02-e0129e99f48b.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eBlotDx: A deep learning tool for Western blot-based diagnostics\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eDeep learning and artificial intelligence (AI), particularly through convolutional neural networks (CNNs) and vision transformers, have revolutionized image-based analysis, offering solutions to automate labor-intensive tasks and standardize results [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. In medical diagnostics, these advancements hold promise for improving accuracy and efficiency. CNNs are now often used in radiology and pathology [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], in combination with transfer learning [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], to help with disease detection, classification, and segmentation. However lab-based diagnostics that involve image interpretation (e.g. Western blot-based assays) have seen relatively little application of deep learning to date.\u003c/p\u003e \u003cp\u003eWestern blot assays, widely employed in clinical diagnostic and research applications, are analytical techniques used to detect specific proteins by analyzing visual band patterns. The process involves separating proteins based on their molecular weight through gel electrophoresis, transferring them onto a membrane, and probing the membrane with antibodies. The presence or absence of antibodies to specific proteins is then visualized as distinct bands, verifying the presence of disease-specific markers or confirming the immune response in a patient sample [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. For example, the Western blot assays used for diagnosing herpes simplex virus (HSV) infection detect the presence of specific antibodies against HSV proteins in a patient's serum [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. These type-specific assays can also determine the subtype of the virus and consist of two separate blots: one with HSV-1 proteins and the second with HSV-2 proteins, resulting in the detection of HSV-1 and/or HSV-2 antibodies in the patient\u0026rsquo;s serum. The ability to distinguish between HSV-1 and HSV-2 antibodies is achieved by the unique protein profiles associated with each virus and accurate interpretation of these bands is critical for diagnostic accuracy.\u003c/p\u003e \u003cp\u003eHSV is a globally prevalent, sexually transmitted virus that causes painful lesions and chronic, lifelong infection. HSV diagnosis can be challenging because lesions may be absent at the time of clinical evaluation, and the frequency of recurrence can be variable across individuals. Serological assays can be used to determine whether the person was previously exposed to HSV even if there are no active lesions present and the results can be important for medical decision-making, communication to partners, family planning, and public health efforts [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Early enzyme-linked immunoassays (ELISAs, [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]) did not distinguish between HSV-1 and HSV-2 and were shown to be inaccurate, particularly in detecting HSV-2 antibodies in HSV-1 seropositive persons [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Type-specific assays based on glycoprotein-G [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] became commercially available in the late 1980s and were shown to be more accurate [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. However, tests based on antibodies to a single protein are more vulnerable to suboptimal sensitivity and specificity [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], as well as to antibody titers and timing, since titers tend to rise 2\u0026ndash;3 months after infection in most cases, and sometimes up to 6 months after infection [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. As a result, HSV Western blot assays continue to be the gold standard due to their ability to more definitively distinguish between HSV-1 and HSV-2, and also to identify seroconversion earlier [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWhen used for clinical testing, Western blot-based assays can be resource-intensive, as they require trained technicians to interpret results manually. The variability in human interpretation can lead to inconsistencies, particularly in ambiguous or borderline cases. Deep learning has demonstrated potential in automating such analyses in other medical imaging applications [\u003cspan additionalcitationids=\"CR18\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. The primary goal is to design a system that achieves the accuracy of trained technicians while minimizing the need for manual interpretation.\u003c/p\u003e \u003cp\u003eThis study introduces convolutional neural network (CNN)-based models for automating the segmentation and classification of Western blot images using HSV serological testing as a use case. The study focuses on automating the analysis of Western blot images by addressing two key objectives. First, the process of cropping blot strips from the original images must be automated to ensure accurate segmentation. Second, the cropped strips must be robustly classified into diagnostic categories. As this was a small pilot study, a relatively simple set of images was used where a diagnosis was obtained after the first round of testing (positive or negative for HSV-1 and HSV-2 separately). Future enhancements will include indeterminate results and edge cases so that BlotDx is capable of handling the full range of variability seen in Western blot images. Common challenges in applying CNNs for Western blot-based diagnostics include variability in image quality, differences in band patterns, and limited availability of annotated datasets. This study aims to overcome these obstacles by integrating robust preprocessing, data augmentation, and model optimization techniques.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDataset: HSV Western blot images and clinical test results\u003c/h2\u003e \u003cp\u003eThe dataset for this study comprised 34 multi-subject images, each containing up to 24 pairs of strips, with each pair corresponding to HSV-1 and HSV-2 serological assay result for a single patient serum sample. The images were generated using a clinically validated HSV Western blot assay (UW-WBA, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://testguide.labmed.uw.edu/view/HSWB\u003c/span\u003e\u003cspan address=\"https://testguide.labmed.uw.edu/view/HSWB\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e, last accessed 23-Mar-25). The UW-WBA is currently the gold-standard method for HSV serology due to its high sensitivity and ability to distinguish between HSV-1 and HSV-2 infection and was performed using methods described previously [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Briefly, HSV-1 and HSV-2 proteins from infected cell lysates were separated by electrophoresis and transferred on to nitrocellulose paper. The resulting strips of paper (blots) were incubated with patient serum and antibodies that bind to the viral proteins were detected by an enzyme-mediated color change. The resulting banding patterns on the blots are read by three independent MLSs, and the consensus used to call the final result as positive, negative, or indeterminate for HSV-1 and HSV-2 antibodies.\u003c/p\u003e \u003cp\u003eBoth type-specific and type-common antibodies to HSV-1 and HSV-2 can be detected in Western blot assays. Type-specific antibodies are directed exclusively against either HSV-1 or HSV-2 antigens, whereas type-common antibodies recognize proteins shared between both virus types [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. For example, a type-common antibody produced in response to an HSV-1 infection may bind to proteins on the HSV-2 blot strip, potentially leading to a false impression of dual infection. As a result, approximately 20% of patient serum samples may require additional steps to clarify the results and accurately determine the infecting virus type. To resolve these ambiguous cases, a pre-adsorption technique is employed. This involves incubating the patient's serum with HSV-1 protein lysate that has been bound to Sepharose beads. During incubation, antibodies specific to HSV-1 as well as type-common antibodies that cross-react with HSV-1 antigens will bind to the lysate and be effectively removed from the sample. The adsorbed serum is then tested using the HSV Western blot assay. Theoretically, only HSV-2\u0026ndash;specific antibodies will remain in the sample, allowing for a clearer interpretation of HSV-2 infection status. Similarly, a parallel adsorption procedure is performed using HSV-2 protein lysate coupled to Sepharose beads. The final result is then called by three MLSs reading the post-adsorption blots. In this pilot study, only blots that had a positive or negative result for HSV-1 and HSV-2 were included, and indeterminate blots and blots needing repeat testing or adsorption were excluded for the classification step. Adsorption blots were only used for the blot identification step (Stage 1 described in the next section).\u003c/p\u003e \u003cp\u003eThe images were divided into 21 for training (n\u0026thinsp;=\u0026thinsp;314 blot pairs), 4 for validation (n\u0026thinsp;=\u0026thinsp;57 blot pairs), and 9 for testing (n\u0026thinsp;=\u0026thinsp;124 blot pairs). In addition, for instance segmentation, one additional image containing 7 groups of 6 strips (from adsorptions) was included in the training set to improve training data diversity (Supplementary Figure S1). Ground truth bounding boxes and masks for training object detection and instance segmentation models were drawn using the VGG Image Annotator [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Ground truth labels, i.e. test results for each strip, were obtained from clinical results in the form of a table containing manual interpretation by 3 technicians, and a final result following supervisory review. The final result was used to create the ground truth labels.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBlotDx: a computer vision and deep learning tool for automating the interpretation of HSV Western blots\u003c/b\u003e \u003c/p\u003e \u003cp\u003eA two-stage framework (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) was adopted to achieve the diagnostic goal of predicting HSV serostatus from input images containing multiple pairs of strips. The first stage crops blot strips from the original images via instance segmentation or object detection methods, and the second stage classifies the cropped strips into diagnostic categories (HSV-1-HSV-2-, HSV-1\u0026thinsp;+\u0026thinsp;HSV-2-, HSV-1-HSV-2+, HSV-1\u0026thinsp;+\u0026thinsp;HSV-2+,),. The two-stage framework was chosen over a one-stage framework (where segmentation/detection and classification are carried out in the same model using the original images as input) because features important for cropping strips may be different from features important for classifying strips.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStage 1: Detecting or segmenting strips using deep learning models\u003c/h3\u003e\n\u003cp\u003eTo detect pairs of strips corresponding to individual subjects, improved Faster R-CNN models with a ResNet-50-FPN backbone were used [\u003cspan additionalcitationids=\"CR24\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] with the number of classes set to 2. To segment individual strips, improved Mask R-CNN models with a ResNet-50-FPN backbone were used with the number of classes set to 2. For both cases, the loss function was adjusted to halve the weight of the classification loss in the total loss computation. Model training was performed by stochastic gradient descent with learning rate 0.001, momentum 0.9, weight decay 0.0005, for 1000 epochs with a batch size of 4. Standard data augmentations including random horizontal and vertical flip, random scaling from 50\u0026ndash;150%, and random rotation of up to 360 degrees were applied. Transfer learning was implemented by starting model training from weights trained on the Common Objects in Context (COCO) dataset.\u003c/p\u003e\n\u003ch3\u003eImage processing for classification\u003c/h3\u003e\n\u003cp\u003eAdditional image processing is needed on the outputs of Stage 1 in order to proceed to classification. The pre-processing required depends on whether instance segmentation or object detection was used for stage 1. If instance segmentation was performed, it outputs masks for the strips which require processing with two main steps: 1) A quadrilateral was fit to each mask. The top edge of the quadrilateral was determined as a horizontal line passing through the 20th percentile of the y-coordinates of the uppermost points in the pair of masks corresponding to the individual containing the input mask. The bottom edge was set as a horizontal line positioned 420 pixels below the top edge. The target rectangle height was chosen such that no bands in the strips were left out. The left edge was estimated by fitting a robust linear regression model to the leftmost points of the mask, while the right edge was estimated similarly using the rightmost points of the mask. 2) The fitted quadrilateral was used to crop out the blot image and transform it into a 23 \u0026times; 420-pixel rectangle using an affine shearing transformation, which preserves the y-coordinate of every point in the trapezoid. The target rectangle width was chosen as the 95th percentile of the strip widths. Once cropped, the two rectangular strips from the same individual were combined either side-by-side to create a 3 x 46 x 420 tensor or stacked top-and-bottom to form a 6 x 23 \u0026times; 420 tensor.\u003c/p\u003e \u003cp\u003eIf object detection was performed in stage 1, it outputs a rectangular bounding box of varying size. A simple pre-processing step is performed to crop the image using the bounding box and rescale the cropped image containing the double strips to be 46 pixels wide and 420 pixels tall.\u003c/p\u003e\n\u003ch3\u003eStage 2: Classification using deep learning models\u003c/h3\u003e\n\u003cp\u003eFor classification, a ResNet-50 model was used to predict HSV-1 and HSV-2 status (positive or negative) separately. Due to cross-reactivity between antibodies against HSV-1 and HSV-2, all classifiers were designed to take both strips as input. Three classifiers were designed (SEG1, SEG2, and DET) varying in how the two strips were provided to the classifier (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Because the strips may be slanted to different degrees in the original image, SEG1 and SEG2 were designed to segment out the two strips separately, transform each into a rectangular shape, and combine them. The relative relationship between the bands on the two strips was maintained during the transformation. The difference between SEG1 and SEG2 lies in how the transformed strips were arranged\u0026mdash;side-by-side in SEG1 and stacked vertically in SEG2. DET was designed as an alternative to SEG1 and SEG2, opting not to \"correct\" the slanting of the strips in the original images. Instead, it employed a bounding box to crop both strips and rescaled the cropped region to a fixed dimension.\u003c/p\u003e \u003cp\u003eModel training was performed by stochastic gradient descent with learning rate 0.001, momentum 0.9, and weight decay 0.0005, for 100 epochs with a batch size of 24. Data augmentations including random scaling from 90\u0026ndash;110% and random rotation of up to 2 degrees were applied. Transfer learning was implemented by starting model training from weights trained on the IMAGENET1K dataset.\u003c/p\u003e\n\u003ch3\u003eClassification model saliency maps\u003c/h3\u003e\n\u003cp\u003eTo gain insights into how the classification models operate, we generate saliency maps [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] using multiple gradient-based methods including class activation maps (CAM) [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], Grad-CAM [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], and Grad-CAM++ [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. These widely used model interpretation methods generate saliency maps by computing weighted combinations of feature maps from convolutional layers of the classification model, which are then up-sampled to the input image resolution using bilinear interpolation. The weights for combining the feature maps reflect each map's contribution to the predicted class and are based on the gradients of the score for a class with respect to the map. For CAM, the gradients are taken with respect to the feature vector after the global average pooling layer. For Grad-CAM, gradients are computed with respect to each pixel in the feature vectors in the final convolutional layer and averaged across the spatial dimensions. Grad-CAM\u0026thinsp;+\u0026thinsp;+\u0026thinsp;generalizes Grad-CAM by replacing simple gradient averaging with a weighted average of the positive gradients. These gradient weights are formally derived by expressing the classification score as a linear combination of the feature vectors after global average pooling.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analyses\u003c/h2\u003e \u003cp\u003eAll statistical analyses were performed using R 4.2.2 (R Project for Statistical Computing).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCode and Data Availability\u003c/h3\u003e\n\u003cp\u003eThe python and R scripts developed for this project are available in the GitHub repository at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/youyifong/BlotDx\u003c/span\u003e\u003cspan address=\"https://github.com/youyifong/BlotDx\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. A snapshot of the repository, taken at the time of this writing, is archived in Zenodo at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.15284041\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.15284041\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The multi-subject images that are the input to Stage 1 contain subject-identifying information and cannot be shared publicly. The single-subject images that are the input to Stage 2 are free of subject-identifying information and available on Zenodo at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.15067008\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.15067008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Classification model checkpoints are available from the same Zenodo repository.\u003c/p\u003e\n\u003ch3\u003eRole of the funding source\u003c/h3\u003e\n\u003cp\u003eNo external funding was received for this study.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003e\u003cstrong\u003eStage 1: Object detection and instance segmentation methods accurately identify blot strips from multi-subject images\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe first stage of our diagnostic tool BlotDx involves identification of blot strips using either object detection or instance segmentation. The two methods were compared in their effectiveness to correctly identify all pairs of blots given an input image with multiple pairs, with each pair containing an HSV-1 (left) strip and HSV-2 (right) strip (Figure 2). The object detection method returns bounding boxes surrounding each pair of strips for a single subject, while the instance segmentation method returns an instance mask, which is converted to a bounding box for individual strips for each subject.\u003c/p\u003e\n\u003cp\u003eFor object detection-based inference, at the minimum confidence score of 0.5 for predicted bounding boxes, all pairs of strips (n=87, 100%) were identified by bounding boxes in the validation set, but many small spurious boxes were also predicted. These were present mainly near the lower boundary of the image possibly due to shadows/artifacts present in the input image (example in Figure S2). Raising the threshold to 0.8 did not remove all the unwanted boxes. Keeping the box confidence score at 0.5 and adding a minimum size for a bounding box to be retained at 1000 pixels removed all unwanted boxes in both the validation set (n = 57 blot pairs) and the test set (n = 124 blot pairs).\u003c/p\u003e\n\u003cp\u003eFor instance segmentation-based inference, using a minimum confidence score of 0.5 for predicted bounding boxes failed to identify all strips in the validation set. Lowering the confidence threshold to 0.1 and introducing a minimum bounding box size of 5000 pixels ensured the detection of all strips without generating unwanted masks. However, multiple masks were occasionally predicted for the same strip. This was addressed by applying an Intersection over Union (IoU) threshold of 0.7 for non-maximum suppression within the Region Proposal Network. Applying these criteria resulted in the detection of all individual strips with no spurious masks in both the validation and test set.\u003c/p\u003e\n\u003cp\u003eThe quality of the inferred masks and bounding boxes was determined by comparing them against ground-truth masks and bounding boxes (GT) for the prediction of serostatus in the validation set (n = 57) using the three classifiers developed (SEG1, SEG2, DET, see Methods and Figure 1). Each classifier was executed three times using different seeds for the random number generator and a majority vote call was determined across the three results. All three classifiers achieved \u0026gt;94% accuracy, and performance was equivalent to or better than classification using GT labels (P-value 0.5, 0.5, 1 for SEG1, SEG2, DET respectively), indicating the adequacy of the inferred masks and bounding boxes for the purposes of classification (Table 1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1.\u0026nbsp;\u003c/strong\u003eClassification accuracy of the three classifiers SEG1, SEG2, and DET using inferred vs. ground truth (GT) masks or bounding boxes based on number of correctly predicted results in the validation set (n = 57 blot pairs) over three replicate runs and an overall number calculated by the majority call.\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/div\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStage 2: Accurate identification of HSV-1 and HSV-2 status using individual and ensemble classifiers\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe performance of the three classifiers using inferred masks and bounding boxes was evaluated in the validation (n = 57) test sets (n = 124), with results from 3 replicate runs combined through majority vote. The three classifiers (SEG1, SEG2, and DET) were evaluated individually and an ensemble approach was also tested that combined the three results through majority voting.\u003c/p\u003e\n\u003cp\u003eFor HSV-1, SEG1 and SEG2 achieved 98.4% and 95.2% accuracy in the test set respectively, while DET achieved perfect classification (Table 2). An ensemble of the three methods achieved accuracies of 98.2% and 99.2% in the validation and test sets respectively. Pooling the results over the validation set and the test sets, the ensemble method achieved a 98.9% accuracy (n = 181). The two misclassified samples were HSV-1 positive misclassified as negative (Figure S3). We repeated the training of classifiers without transfer learning, i.e. starting from random weights instead of ImageNet-trained weights. For the ensemble method, forgoing transfer learning led to 5 additional HSV-1 positive samples misclassified as negative with no gains. The McNemar\u0026apos;s test comparing results with and without transfer learning yielded a P-value of 0.0625, suggesting that models benefited from pre-trained weights.\u003c/p\u003e\n\u003cp\u003eFor HSV-2, the ensemble method achieved 97.8% (177/181) accuracy overall (97.2% for SEG1, 97.8% for SEG2, 97.2% for DET, Table 2). The four misclassified samples included two that were HSV-2 positive misclassified as negative and two that were misclassified in the opposite direction. Starting model training from random weights led to the level of accuracy for the ensemble method.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u0026nbsp;\u003c/strong\u003eClassification accuracy of the three classifiers SEG1, SEG2, and DET on the test set (n = 124) and a combined set of validation and test images (n = 181) with and without transfer learning (TL).\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003eWith TL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003eNo TL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003eTest\u0026nbsp;\u003cbr\u003e\u0026nbsp;(n = 124)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003eCombined\u0026nbsp;\u003cbr\u003e\u0026nbsp;(n = 181)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003eCombined\u0026nbsp;\u003cbr\u003e\u0026nbsp;(n = 181)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eP-value*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 78px;\"\u003e\n \u003cp\u003eSEG1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e122\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e178\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e175\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e0.25\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e119\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e176\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e177\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 78px;\"\u003e\n \u003cp\u003eSEG2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e118\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e177\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e168\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e120\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e177\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e173\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 78px;\"\u003e\n \u003cp\u003eDET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e124\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e178\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e174\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e119\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e176\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e177\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 78px;\"\u003e\n \u003cp\u003eEnsemble\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e123\u003c/p\u003e\n \u003cp\u003e99.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e179\u003c/p\u003e\n \u003cp\u003e98.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e174\u003c/p\u003e\n \u003cp\u003e96.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003eHSV-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e120\u003c/p\u003e\n \u003cp\u003e96.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e177\u003c/p\u003e\n \u003cp\u003e97.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e177\u003c/p\u003e\n \u003cp\u003e97.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 54px;\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u0026nbsp;*McNemar\u0026rsquo;s test\u003c/p\u003e\n\u003cp\u003eOverall, when implemented for both subtypes with transfer learning and using an ensemble method, BlotDx produced 98.9% test accuracy of, 97.6% sensitivity, and 100% specificity for HSV-1, and 97.8% test accuracy, 95.5% sensitivity, and 98.5% specificity for HSV-2 (Table 3).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u0026nbsp;\u003c/strong\u003eConfusion matrices for ensemble results with transfer learning\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 103px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 173px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHSV-1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 173px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHSV-2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 103px;\"\u003e\n \u003cp\u003eTrue\\Predicted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003eNegative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003ePositive\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003eNegative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003ePositive\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 103px;\"\u003e\n \u003cp\u003eNegative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e134\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 103px;\"\u003e\n \u003cp\u003ePositive\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 87px;\"\u003e\n \u003cp\u003e43\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u0026nbsp;To obtain a more robust estimate of classification performance, we conducted a 5-fold cross-validation experiment. All blot pairs were randomly partitioned into five folds. Models were trained on four folds and evaluated on the remaining fold. The ensemble approach achieved classification accuracies of 98.8% for HSV-1 and 98.6% for HSV-2 (Table S1). The slight performance improvement may be attributed to the higher training-to-test data ratio in the 5-fold cross-validation setup.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePaired HSV-1 and HSV-2 blots provide improved classification accuracy compared to single blots\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the models above, due to expected cross-reactivity between HSV-1 and HSV-2 antigens (typically referred to as type-common antibodies), all classifiers were designed to take HSV-1 and HSV-2 strips for each subject together as input. To assess the impact of cross-reactivity on classification accuracy, these paired-strip classifiers were compared against an alternative single-strip design. In the single strip design, the HSV-1 result was generated by a classifier that only received the HSV-1 strip as input, and likewise for the HSV-2 classifier. The single-strip classifiers achieved a lower overall classification accuracy (93.4% for HSV-1 and 97.2% for HSV-2, Table 4), suggesting that information from both blots is important for classification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4.\u0026nbsp;\u003c/strong\u003eClassification accuracy of a single-strip classifier using SEG1 on the validation set (n = 57) and test set images (n = 124) with and without transfer learning (TL).\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003eWith TL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003eNo TL\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003eValidation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003eValidation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003eHSV-1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003e52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e117\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e111\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003eHSV-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003e57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e119\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e119\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003eClassification model saliency maps\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSaliency maps were used to visualize the relative importance of different spatial regions to the classification result. Four different algorithms were compared (CAM, GradCAM, and GradCAM++, see Methods). Comparing the HSV-1 and HSV-2 saliency maps (Figure 3), we see that the areas that contribute to HSV-2 positive identification are limited to the HSV-2 strip, while the areas that contribute to HSV-1 positive identification appear to spread between both the HSV-1 strip and the HSV-2 strip. This supports our previous result that showed that both blot strips are required for accurate classification. The saliency maps from the HSV-1 models also show greater image-to-image variability than those from the HSV-2 model.\u003c/p\u003e\n\u003cp\u003eThe four different algorithms had good visual correspondence with each other, but there were some differences in the importance regions for SEG1 and DET classifiers, particularly for HSV-1 models (Figure 3A). For the prediction of HSV-1+/HSV-2- status, high significance regions for HSV-1 models (Figure 3A(i)) included bands corresponding to HSV-2 glycoproteins gB/gG, gC, gE, and major capsid protein VP5 [30,31] for both SEG1 and DET classifiers. The assembly protein ICP35 region was highlighted in both HSV-1 and HSV-2 blots for the SEG1 classifier, but only on the HSV-2 blot for the DET classifier. For the prediction of HSV-1+/HSV-2+ status by the HSV-1 models (Figure 3A(iii)), the SEG1 classifier focused primarily on the gB/gG region of the HSV-1 blot, while the DET classifier focused on the gD region of the HSV-2 blot. The HSV-2 models showed fewer differences between SEG1 and DET classifiers and in all cases the highlighted regions of high importance for both single and double-positive cases (HSV-1-/HSV-2+ and HSV-1+/HSV-2+ in Figure 3B) we on the HSV-2 blot alone. Of note, both SEG1 and DET classifiers for HSV-2 models strongly highlighted the gB/gG region, which is consistent with the way the blots are interpreted in the lab where a strong gG band on the HSV-2 blot is a key requirement for an HSV-2 positive result.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThis study presents an AI-based tool (BlotDx) for the analysis of diagnostic Western blots. While AI-based methods have been applied to Western blots in other settings like plagiarism detection [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], BlotDx is, to our knowledge, the first such application of CNNs for serodiagnostics in a clinical laboratory setting. We present a two-stage deep learning solution that performs identification of individual or paired blot strips from multiple-sample input images and then performs classification to return a result for HSV-1 and HSV-2 serostatus. The proposed two-stage solution identified all strips from multi-subject images and achieved high diagnostic accuracy for HSV-1 and HSV-2. Given that the current laboratory workflow involves interpretation by multiple technicians followed by final supervisory review, BlotDx offers opportunities for cost savings and efficiency by potentially simplifying the process to require fewer reviewers. Although currently designed for HSV serological assays, the same approach could be easily adapted for other applications that employ Western blots, e.g. for Lyme disease [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e], human T-lymphotrophic virus [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e], bovine spongiform encephalopathy [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e], and tapeworm infections like Echinococcus [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe findings also highlight the value of transfer learning and ensemble methods, and of combining object detection and instance segmentation models in a two-stage process for improved model performance and stability. While modern computer vision models can simultaneously identify and classify objects, two-stage approaches that separate identification from classification have demonstrated improved performance in various practical applications, such as cancer diagnosis [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], skin lesion analysis [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e], and splenomegaly classification [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAs a step towards explainable AI, we used saliency maps to identify regions of importance within the input image, highlighting which pixels or regions contributed the most towards a classification result. The finding that the HSV-1 model\u0026rsquo;s saliency maps encompass signal from both HSV-1 and HSV-2 blots may be related to our finding that single-strip classifiers underperform compared to classifiers that use both strips as input. One possibility is that the model may have implicitly learned that information from the HSV-2 blot is predictive of the HSV-1 label\u0026mdash;either due to biological correlation or experimental artifact\u0026mdash;and thus relies on both regions for its decision-making. The second explanation is that this phenomenon reflects an instance of aliasing as described by Li et al.[\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], where spatially distinct but statistically correlated features introduce misleading attributions in saliency maps. This could occur if regions within the HSV-1 and HSV-2 blots exhibit high inter-feature correlation (perhaps in a subset of positive training samples), causing the gradients to erroneously attribute relevance across both regions. The third possibility is that the saliency maps, derived from gradients of the final feature extraction layer, represent higher-level semantics and fail to localize the spatial source of discriminative signal accurately. This explanation, however, appears less compelling given that the saliency maps generated from the HSV-2 model are confined to the HSV-2 blot, suggesting that spatial localization is at least partially preserved in the saliency pipeline.\u003c/p\u003e \u003cp\u003eResults from saliency maps are also consistent with the way the blots are read in the lab. MLSs interpreting the blots use information from both strips, and particular attention is paid to glycoprotein G which, when present as a strong band on the HSV-2 blot indicates a high probability of HSV-2 infection. HSV-1 and HSV-2 are closely related herpesviruses with \u0026gt;\u0026thinsp;70% sequence identity, reflecting their shared evolutionary origin [\u003cspan additionalcitationids=\"CR43 CR44\" citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. This relatedness gives rise to both type-common and type-specific antigens. Western blot assays detect a spectrum of antibody responses, producing a mixture of shared and unique banding patterns. Key type-specific antigens\u0026mdash;such as glycoprotein G-1 (gG-1) for HSV-1 and glycoprotein G-2 (gG-2) for HSV-2\u0026mdash;enable differentiation between the two subtypes when interpreted correctly. However, reliance on gG alone has shown to be less sensitive to detecting seropositivity when compared to using the full spectrum of bands in the UW-WBA [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWhile this work demonstrates the potential utility of AI in lab-based diagnostics, our study had a number of limitations. As this was a pilot study, a relatively small dataset was used and strips with indeterminate results that required re-testing or adsorption were excluded. Data augmentation was used to mitigate the small size of the dataset to some extent but a larger dataset with greater complexity, and classification with more result classes would improve the usability in a practical setting. As the images were obtained by photographing archived blot strips, some fading is expected over time, and this may have contributed to some of the misclassifications. The use of freshly run blots, and photographing strips on a dark background to enhance contrast may lead to further improvement in classification accuracy. The black-box nature of CNNs raises concerns about interpretability, and future work should aim to integrate tools that provide insights into model decisions to enhance trust among clinicians. There is also potential for incorporating the image capture and analysis into an automated benchtop solution that accelerates the sample-to-answer process. Overall, this work paves the way for AI-powered solutions that bridge the gap between traditional diagnostics and modern computational methods.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eLeCun Y., Bengio Y., Hinton G. Deep learning. Nature. 2015;521: 436\u0026ndash;444. PMID 26017442\u003c/li\u003e\n\u003cli\u003eDosovitskiy A., Beyer L., Kolesnikov A., Weissenborn D., Zhai X., Unterthiner T., Dehghani M., Minderer M., Heigold G., Gelly S., Uszkoreit J., Houlsby N. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv [cs.CV]. 2020. Available: http://arxiv.org/abs/2010.11929\u003c/li\u003e\n\u003cli\u003eLitjens G., Kooi T., Bejnordi B.E., Setio A.A.A., Ciompi F., Ghafoorian M., van der Laak J.A.W.M., van Ginneken B., S\u0026aacute;nchez C.I. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42: 60\u0026ndash;88. PMID 28778026\u003c/li\u003e\n\u003cli\u003eEsteva A., Chou K., Yeung S., Naik N., Madani A., Mottaghi A., Liu Y., Topol E., Dean J., Socher R. Deep learning-enabled medical computer vision. NPJ Digit Med. 2021;4: 5. PMID 33420381PMCID PMC7794558\u003c/li\u003e\n\u003cli\u003eKim H.E., Cosa-Linan A., Santhanam N., Jannesari M., Maros M.E., Ganslandt T. Transfer learning for medical image classification: a literature review. BMC Med Imaging. 2022;22: 69. PMID 35418051PMCID PMC9007400\u003c/li\u003e\n\u003cli\u003eTowbin H., Staehelin T., Gordon J. Electrophoretic transfer of proteins from polyacrylamide gels to nitrocellulose sheets: procedure and some applications. Proc Natl Acad Sci U S A. 1979;76: 4350\u0026ndash;4354. PMID 388439PMCID PMC411572\u003c/li\u003e\n\u003cli\u003eBegum H., Murugesan P., Tangutur A.D. Western blotting: a powerful staple in scientific and biomedical research. Biotechniques. 2022;73: 58\u0026ndash;69. PMID 35775367\u003c/li\u003e\n\u003cli\u003eWorkowski K.A., Bachmann L.H., Chan P.A., Johnston C.M., Muzny C.A., Park I., Reno H., Zenilman J.M., Bolan G.A. Sexually transmitted infections treatment guidelines, 2021. MMWR Recomm Rep. 2021;70: 1\u0026ndash;187. PMID 34292926PMCID PMC8344968\u003c/li\u003e\n\u003cli\u003eKimmel N., Friedman M.G., Sarov I. Enzyme-linked immunosorbent assay (ELISA) for detection of herpes simplex virus-specific IgM antibodies. J Virol Methods. 1982;4: 219\u0026ndash;227. PMID 6286702\u003c/li\u003e\n\u003cli\u003eAshley R., Cent A., Maggs V., Nahmias A., Corey L. Inability of enzyme immunoassays to discriminate between infections with herpes simplex virus types 1 and 2. Ann Intern Med. 1991;115: 520\u0026ndash;526. PMID 1652909\u003c/li\u003e\n\u003cli\u003eWald A., Ashley-Morrow R. Serological testing for herpes simplex virus (HSV)-1 and HSV-2 infection. Clin Infect Dis. 2002;35: S173-82. PMID 12353203\u003c/li\u003e\n\u003cli\u003eKatz D., Hilliard J.K., Mirkovic R.R., Word R.A. ELISA for detection of IgG and IgM antibodies to HSV-1 and HSV-2 in human sera. J Virol Methods. 1986;14: 43\u0026ndash;55. PMID 3021796\u003c/li\u003e\n\u003cli\u003eLee F.K., Coleman R.M., Pereira L., Bailey P.D., Tatsuno M., Nahmias A.J. Detection of herpes simplex virus type 2-specific antibody with glycoprotein G. J Clin Microbiol. 1985;22: 641\u0026ndash;644. PMID 3001136PMCID PMC268483\u003c/li\u003e\n\u003cli\u003eAshley R.L. Sorting out the new HSV type specific antibody tests. Sex Transm Infect. 2001;77: 232\u0026ndash;237. PMID 11463920PMCID PMC1744343\u003c/li\u003e\n\u003cli\u003eCrawford K.H.D., Selke S., Pepper G., Goecker E., Sobel A., Wald A., Johnston C., Greninger A.L. Performance characteristics of highly automated HSV-1 and HSV-2 IgG testing. J Clin Microbiol. 2024;62: e0026324. PMID 38687020PMCID PMC11237592\u003c/li\u003e\n\u003cli\u003eAshley R.L., Wald A. Genital herpes: review of the epidemic and potential use of type-specific serology. Clin Microbiol Rev. 1999;12: 1\u0026ndash;8. PMID 9880471PMCID PMC88903\u003c/li\u003e\n\u003cli\u003ePinto-Coelho L. How artificial intelligence is shaping medical imaging technology: A survey of innovations and applications. Bioengineering (Basel). 2023;10. PMID 38136026PMCID PMC10740686doi: 10.3390/bioengineering10121435\u003c/li\u003e\n\u003cli\u003eLiu X., Faes L., Kale A.U., Wagner S.K., Fu D.J., Bruynseels A., Mahendiran T., Moraes G., Shamdas M., Kern C., Ledsam J.R., Schmid M.K., Balaskas K., Topol E.J., Bachmann L.M., Keane P.A., Denniston A.K. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health. 2019;1: e271\u0026ndash;e297. PMID 33323251\u003c/li\u003e\n\u003cli\u003eVaroquaux G., Cheplygina V. Machine learning for medical imaging: methodological failures and recommendations for the future. NPJ Digit Med. 2022;5: 48. PMID 35413988PMCID PMC9005663\u003c/li\u003e\n\u003cli\u003eAshley R.L., Militoni J., Lee F., Nahmias A., Corey L. Comparison of Western blot (immunoblot) and glycoprotein G-specific immunodot enzyme assay for detecting antibodies to herpes simplex virus types 1 and 2 in human sera. J Clin Microbiol. 1988;26: 662\u0026ndash;667. PMID 2835389PMCID PMC266403\u003c/li\u003e\n\u003cli\u003eBernstein D.I., Bryson Y.J., Lovett M.A. Antibody response to type-common and type-unique epitopes of herpes simplex virus polypeptides. J Med Virol. 1985;15: 251\u0026ndash;263. PMID 2580054\u003c/li\u003e\n\u003cli\u003eDutta A., Zisserman A. The VIA Annotation Software for Images, Audio and Video. Proceedings of the 27th ACM International Conference on Multimedia. New York, NY, USA: ACM; 2019. doi: 10.1145/3343031.3350535\u003c/li\u003e\n\u003cli\u003eLi Y., Xie S., Chen X., Dollar P., He K., Girshick R. Benchmarking detection transfer learning with vision transformers. arXiv [cs.CV]. 2021. Available: http://arxiv.org/abs/2111.11429\u003c/li\u003e\n\u003cli\u003eRen S., He K., Girshick R., Sun J. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans Pattern Anal Mach Intell. 2017;39: 1137\u0026ndash;1149. PMID 27295650\u003c/li\u003e\n\u003cli\u003eHe K., Gkioxari G., Dollar P., Girshick R. Mask R-CNN. 2017 IEEE International Conference on Computer Vision (ICCV). IEEE; 2017. pp. 2980\u0026ndash;2988.\u003c/li\u003e\n\u003cli\u003eSzczepankiewicz K., Popowicz A., Charkiewicz K., Nałęcz-Charkiewicz K., Szczepankiewicz M., Lasota S., Zawistowski P., Radlak K. Ground truth based comparison of saliency maps algorithms. Sci Rep. 2023;13: 16887. PMID 37803108PMCID PMC10558518\u003c/li\u003e\n\u003cli\u003eZhou B., Khosla A., Lapedriza A., Oliva A., Torralba A. Learning deep features for discriminative localization. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2016. doi: 10.1109/cvpr.2016.319\u003c/li\u003e\n\u003cli\u003eSelvaraju R.R., Cogswell M., Das A., Vedantam R., Parikh D., Batra D. Grad-CAM: Visual explanations from deep networks via gradient-based localization. Int J Comput Vis. 2020;128: 336\u0026ndash;359.\u003c/li\u003e\n\u003cli\u003eChattopadhay A., Sarkar A., Howlader P., Balasubramanian V.N. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE; 2018. doi: 10.1109/wacv.2018.00097\u003c/li\u003e\n\u003cli\u003eMcKendall R.R., Pettit M., Woo W. The polyspecific immunoglobulin response to HSV-1 viral proteins: determination of immunogenic proteins and relative antibody titres to individual polypeptides by immunoblotting. J Med Microbiol. 1988;25: 49\u0026ndash;57. PMID 2826789\u003c/li\u003e\n\u003cli\u003eWales S.Q., Smith C.C., Wachsman M., Calton G., Aurelian L. Performance and use of a ribonucleotide reductase herpes simplex virus type-specific serological assay. Clin Diagn Lab Immunol. 2004;11: 42\u0026ndash;49. PMID 14715543PMCID PMC321330\u003c/li\u003e\n\u003cli\u003eGosselin R.-D. AI detectors are poor western blot classifiers: a study of accuracy and predictive values. PeerJ. 2025;13: e18988. PMID 39989748PMCID PMC11847483\u003c/li\u003e\n\u003cli\u003eDressler F., Whalen J.A., Reinhardt B.N., Steere A.C. Western blotting in the serodiagnosis of Lyme disease. J Infect Dis. 1993;167: 392\u0026ndash;400. PMID 8380611\u003c/li\u003e\n\u003cli\u003eJohn T.M., Taege A.J. Appropriate laboratory testing in Lyme disease. Cleve Clin J Med. 2019;86: 751\u0026ndash;759. PMID 31710588\u003c/li\u003e\n\u003cli\u003eRibeiro L.C.P., Gon\u0026ccedil;alves C.C.A., Slater C.M.S.A., Carvalho S.M.F. de, Puccioni-Sohler M. Standardisation of Western blotting to detect HTLV-1 antibodies synthesised in the central nervous system of HAM/TSP patients. Mem Inst Oswaldo Cruz. 2013;108: 730\u0026ndash;734. PMID 24037195PMCID PMC3970696\u003c/li\u003e\n\u003cli\u003eOkada H., Murayama Y., Shimozaki N., Yoshioka M., Masujin K., Imamura M., Iwamaru Y., Matsuura Y., Miyazawa K., Fukuda S., Yokoyama T., Mohri S. Prion in saliva of bovine spongiform encephalopathy\u0026ndash;infected cattle. Emerg Infect Dis. 2012;18: 2091\u0026ndash;2092.\u003c/li\u003e\n\u003cli\u003eDeininger S., Wellinghausen N. Evaluation of a new combined Western and line blot assay (EUROLINE-WB) for diagnosis and species identification of Echinococcus infection in humans. GMS Infect Dis. 2019;7: Doc01. PMID 30993059PMCID PMC6446614\u003c/li\u003e\n\u003cli\u003eMadhu G., Meher Bonasi A., Kautish S., Almazyad A.S., Mohamed A.W., Werner F., Hosseinzadeh M., Shokouhifar M. UCapsNet: A two-stage deep learning model using U-Net and Capsule Network for breast cancer segmentation and classification in ultrasound imaging. Cancers (Basel). 2024;16. PMID 39594732PMCID PMC11592068doi: 10.3390/cancers16223777\u003c/li\u003e\n\u003cli\u003eInnani S., Dutande P., Baheti B., Baid U., Talbar S. Deep learning based novel cascaded approach for skin lesion analysis. InInternational Conference on Computer Vision and Image Processing. Cham; Nature Switzerland: Springer; 2022. pp. 615\u0026ndash;626.\u003c/li\u003e\n\u003cli\u003eMeddeb A., Kossen T., Bressem K.K., Molinski N., Hamm B., Nagel S.N. Two-stage deep learning model for automated segmentation and classification of splenomegaly. Cancers (Basel). 2022;14: 5476. PMID 36428569PMCID PMC9688308\u003c/li\u003e\n\u003cli\u003eLi K., Wu Z., Peng K.-C., Ernst J., Fu Y. Tell me where to look: Guided attention inference network. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE; 2018. doi: 10.1109/cvpr.2018.00960\u003c/li\u003e\n\u003cli\u003eAguilar J.S., Devi-Rao G.V., Rice M.K., Sunabe J., Ghazal P., Wagner E.K. Quantitative comparison of the HSV-1 and HSV-2 transcriptomes using DNA microarray analysis. Virology. 2006;348: 233\u0026ndash;241. PMID 16448680\u003c/li\u003e\n\u003cli\u003eBaines J.D., Pellett P.E. Genetic comparison of human alphaherpesvirus genomes. In: Arvin A, Campadelli-Fiume G, Mocarski E, Moore PS, Roizman B, Whitley R, et al., editors. Human Herpesviruses. Cambridge: Cambridge University Press; 2007. pp. 61\u0026ndash;69.\u003c/li\u003e\n\u003cli\u003eSchmid D.S. Mixing it up: New insights into interspecies recombination between herpes simplex virus type 1 and 2. The journal of infectious diseases. Oxford University Press (OUP); 2020. pp. 1208\u0026ndash;1209. PMID 31107962PMCID PMC9137430\u003c/li\u003e\n\u003cli\u003eCasto A.M., Roychoudhury P., Xie H., Selke S., Perchetti G.A., Wofford H., Huang M.-L., Verjans G.M.G.M., Gottlieb G.S., Wald A., Jerome K.R., Koelle D.M., Johnston C., Greninger A.L. Large, Stable, Contemporary Interspecies Recombination Events in Circulating Human Herpes Simplex Viruses. J Infect Dis. 2020;221: 1271\u0026ndash;1279. PMID 31016321PMCID PMC7325804\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University of Washington","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"deep learning, Western blots, herpes simplex virus, diagnostics, clinical testing, serology","lastPublishedDoi":"10.21203/rs.3.rs-6605358/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6605358/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWestern blot (WB) is the gold standard for herpes simplex virus (HSV-1/HSV-2) serology but requires manual interpretation by trained laboratory personnel, which is time-consuming and variable. Automating this process could improve efficiency and diagnostic consistency. This study presents BlotDx, a deep learning tool for automated interpretation of HSV WB assays.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBlotDx uses a two-stage approach: (1) instance segmentation or object detection to identify blot strips from input images, and (2) classification models to determine HSV-1 and HSV-2 serostatus. The dataset used contained 34 images, each with up to 24 pairs of blot strips. Of these, 21 images were used for training, 4 for validation, and 9 for testing (n = 314, 57, and 124 blot pairs respectively). Data augmentation, ensemble modeling, and transfer learning were used to enhance model performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFindings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCompared to the gold standard of expert human review by three independent MLSs, BlotDx demonstrated high diagnostic accuracy (99% for HSV-1, 98% for HSV-2) across validation and test datasets. Data augmentation and ensemble methods significantly improved robustness and reliability. The transfer learning approach allowed the models to leverage pre-trained weights, further enhancing performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInterpretation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study highlights the utility of AI in automating the analysis of Western blots for HSV diagnostics, with potential applications in other diseases. The proposed two-stage approach, combined with modern deep learning techniques, achieved scalable, accurate, and efficient diagnostic results. These findings underscore the potential of deep learning to transform traditional diagnostic workflows by reducing costs and increasing efficiency.\u003c/p\u003e","manuscriptTitle":"BlotDx: A deep learning tool for Western blot-based diagnostics","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-09 11:26:07","doi":"10.21203/rs.3.rs-6605358/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"bbd46566-f960-404b-9a01-133c0ba26aa9","owner":[],"postedDate":"May 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-05-09T11:26:07+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-09 11:26:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6605358","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6605358","identity":"rs-6605358","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}