Peering Inside the Black Box: Explainable AI to Interpret Advanced Computer Vision Fungal Pathogen Prediction

Peering Inside the Black Box: Explainable AI to Interpret Advanced Computer Vision Fungal Pathogen Prediction | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (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];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Peering Inside the Black Box: Explainable AI to Interpret Advanced Computer Vision Fungal Pathogen Prediction View ORCID Profile Joshua D. Guthrie , View ORCID Profile Shamanth A. Shankarnarayan , View ORCID Profile Daniel A. Charlebois doi: https://doi.org/10.1101/2025.06.27.662051 Joshua D. Guthrie 1 Department of Physics, University of Alberta , Edmonton, Alberta, Canada Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Joshua D. Guthrie Shamanth A. Shankarnarayan 1 Department of Physics, University of Alberta , Edmonton, Alberta, Canada Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Shamanth A. Shankarnarayan Daniel A. Charlebois 1 Department of Physics, University of Alberta , Edmonton, Alberta, Canada 2 Department of Biomedical Engineering, University of Alberta , Edmonton, Alberta, Canada 3 Department of Biological Sciences, University of Alberta , Edmonton, Alberta, Canada Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Daniel A. Charlebois For correspondence: dcharleb{at}ualberta.ca Abstract Full Text Info/History Metrics Supplementary material Preview PDF ABSTRACT Antimicrobial resistance is a growing concern, with pathogenic fungi making a substantial contribution to untreatable life-threatening infections across the globe. Artificial intelligence is increasingly used in microbiology and antimicrobial resistance research, with promise to improve clinical diagnostics and infectious disease treatments. However, how AI models make predictions remains largely unknown, which is a major hurdle for human trust and regulatory approval. We trained convolutional neural networks (DenseNet121 and InceptionV3) and vision transformers (Swin Transformer-Tiny and Vision Transformer-Base 16) to quickly and accurately identify human fungal pathogens from microscopy images. Using explainable AI (Occlusion Sensitivity and Grad-CAM), we identified biologically relevant features (organelle, cell interior, cell wall, budding patterns/scars, and optical patterns) and irrelevant image features (background artifacts) high-performance computer vision models used to make predictions. These findings advance our understanding of how computer vision models make predictions on microbial pathogens and are anticipated to have profound implications for AI-based diagnostics. 1 Introduction Globally, 6.5 million fungal infections are estimated to occur annually and are associated with 3.8 million deaths, of which 2.5 million deaths are directly attributed to these infections 1 . Candida bloodstream infection (candidemia) and invasive candidiasis result in nearly 1 million deaths annually 1 . Many fungal infections are underreported and underdiagnosed, and treatment is restricted to a few classes of antifungal drugs 2 . Candidemia is the fourth most common infectious disease in intensive care units and occurs predominantly in hospitalized patients with underlying risk factors; however, candidemia is not limited to patients who are immune-compromised and/or critically ill 3 . The emergence of multiple new human fungal pathogens each year 2 and increasing resistance to antifungal drugs has escalated the demands on healthcare facilities, leading to higher economic burden and mortality 4 . Candida albicans is the leading cause of candidemia in many geographical locations around the world. However, non-albicans species such as Nakseomyces glabratus (previously Candida glabrata ), Pichia kudriavzevii (previously C. krusei ), Candida parapsilosis , and Candida tropicalis contribute up to 95% of the disease burden 3 , 5 . The multidrug-resistant pathogenic yeast Candidozyma auris (previously Candida auris ) has appeared in over 60 countries over the last decade, with hospital-acquired infections and outbreaks being a major concern 6 . Multiple challenges exist to identify pathogenic yeast species. The blood culture method is the gold standard for identifying yeasts in candidemia cases 7 . Alternatively, non-culture-based methods, such as identifying serum biomarkers ( β -(1,3)-D-glucan) and nucleic acids (nucleic acid amplification tests) 8 , are also employed. However, these methods are limited due to low sensitivity, long turnaround times, and sophisticated instrumentation, specialized laboratory space, and highly-trained personnel requirements 7 . To overcome these challenges, several new detection methods are being developed for faster detection of yeast species directly from blood samples, as well as yeasts grown on agar plates 7 . These detection methods include the use of chrome agar 9 , matrix-assisted laser desorption/ionization-time of flight mass spectroscopy (MALDI TOF MS) 10 , DNA sequencing 11 , T2 magnetic resonance 12 , volatile organic compounds 13 , Fourier transform infrared with attenuated total reflectance 14 , imperfect match probes 15 , recombinase polymerase amplification 16 , high resolution melting curves 17 , Raman spectroscopy 18 , and laser-induced breakdown spectroscopy 19 . However, many challenges exist, such as integrating these methods into routine diagnostic workflows, species-level misidentification, and limited access in low-and middle-income countries 1 . The use of artificial intelligence (AI) and machine learning (ML) has seen remarkable growth in recent years, driven by advances in model architectures, training algorithms, computational hardware, open-source AI frameworks (e.g., PyTorch 20 ), and the availability of vast amounts of training data 21 . Artificial intelligence and ML approaches are increasingly being applied in scientific research to model complex data 21 . For instance, AlphaFold can determine the 3D structures of proteins and their bimolecular interactions with accuracies comparable to experiments, and is accelerating drug discovery 22 . Recent studies have investigated ML methods in the diagnosis and treatment of infections 23 , and to discover antibiotics 24 . The field of computer vision, which applies AI to analyze image data, has made substantial progress from advances in artificial neural networks and attention-based transformer models 25 – 27 . Computer vision models based on convolutional neural network (CNN) and vision transformer architectures have achieved state-of-the-art performance on gold-standard image classification tasks 25 – 28 , such as the ImageNet competition 26 . These models have been applied in biological and medical image analysis 29 , achieving performance comparable to current diagnostic tools and human professionals 30 . They have also been applied in recent studies for the classification of yeast species in microscopy images 31 – 33 . Furthermore, advances in computer vision segmentation models have enabled advanced automated image processing 34 , finding various applications in cell microscopy image analysis 35 – 38 . Deep learning model architectures have been developed for computer vision applications. Deep learning methods have been applied in the context of yeast pathogens, including segmenting microscopy images 37,39 and inner organelles 40 , tracking cellular growth 41 , and measuring cell dimensions 42 . InceptionV3 43 , a deep CNN architecture with 48 layers and “inception modules” that apply convolutions at various filter sizes to capture features at multiple scales, was previously combined with transfer learning to classify Candida species 33 . Specifically, the InceptionV3-based model was able to classify four pathogenic yeast species ( C. albicans, C. auris, N. glabratus , and C. haemulonii ) from brightfield microscopy images with an overall accuracy of 78%. DenseNet121, ViT, and Swin Transformer models have not been used for yeast species classification and provide improvements over InceptionV3 and other computer vision architectures 25 , 44 – 46 . DenseNet121 is a deep CNN architecture with 121 layers that features a dense connectivity pattern that connects each layer in the network to all preceding layers 44 . The DenseNet architecture was designed to improve gradient calculations and promote the reuse of features, leading to better efficiency through faster optimization and smaller network and dataset size requirements to reach high performance 44 . Vision transformers are relatively new models that depart from the CNN architecture by adapting the transformer architecture of large language models to computer vision tasks, using self-attention mechanisms to better learn global features in images 45 , resulting in comparable or better results than CNNs on standard image classification metrics 45 . However, vision transformers require more data than CNNs to reach comparable performance 25 . Swin (Shifted Window) transformers are a modification of the vision transformer architecture that addresses this limitation through a hierarchical representation of feature maps and the use of shifting window attention, improving local feature extraction capabilities while reducing computational complexity 46 . The “black box” nature of AI models often renders their predictions difficult, or even impossible, to understand. This limits the application of AI in fields where model interpretation and human trust are of particular importance, such as scientific research and medical diagnostics 47 . The emerging field of explainable AI (XAI) aims to elucidate how AI/ML models make predictions 48 , with an emphasis on computer vision 49 . Although complete explainability of computer vision models has yet to be achieved, XAI methods can identify image features that are most important for model predictions 49 . Understanding the decision-making process underlying AI models will enable scientific and medical fields to adopt AI-driven tools with greater confidence. Explainable AI methods investigate predictions made by computer vision models, such as Grad-CAM (Gradient Class Activation Mapping) 50 and Occlusion Sensitivity 51 . Grad-CAM uses information flowing into the final convolutional layer of a CNN to produce a coarse-grained visualization of areas in an image that were most important for a given prediction 50 . This makes Grad-CAM suitable for determining if the model places high prediction importance on a high-level object within an image, but not for determining specific predictive features. Occlusion Sensitivity is a model-agnostic perturbation method that measures the change in prediction output of a model when occluding specific regions of the input image, with large changes in model output indicating prediction importance of the occluded region 51 . Occlusion Sensitivity is capable of producing fine-grained importance visualizations to identify specific features that lead to model predictions 51 . In this study, we develop high-performance computer vision models (InceptionV3 43 , DenseNet121 44 , Vision Transformer (ViT)-Base 16) 45 , and Swin Transformer-Tiny 46 ) to quickly and accurately identify yeast pathogens from microscopy images. We overcome current limitations in computer vision pathogen identification by applying XAI to interpret the predictions made by these advanced computer vision classification models. We use segmentation models to automate data processing, enabling the generation of large datasets for model training and evaluation, while increasing data processing efficiency, eliminating a major bottleneck of previous work 31 – 33 . This increase in training data and improvements in model development, particularly the use of state-of-the-art model architectures and full model fine-tuning, leads to substantial improvements in model performance that scales with classification complexity through the inclusion of additional pathogenic yeast species (e.g., P. kudriavzevii, C. parapsilosis , and C. tropicalis ). The use of XAI methods to interpret computer vision predictions enables model verification, determines biologically relevant and irrelevant features used for species predictions, identifies areas for model improvement, and enhances trust in the application of AI models for pathogen identification. 2 Materials and Methods 2.1 Sample Preparation, Imaging, and Data Processing Clinical isolates of C. albicans, C. auris , and N. glabratus were obtained from the Alberta Precision Laboratory -Public Health Laboratory. the The American Type Culture Collection (ATCC) provided the strains for C. parapsilosis ATCC-22019, P. kudriavzevii ATCC-6258, C. haemulonii ATCC-22991, and C. tropicalis ATCC-750. All strains were preserved in 25% glycerol at -80°C until further use. The clinical and standard strains were revived by culturing from frozen stock in Sabouraud glucose agar (SDA) plates (Millipore, Darmstadt, Germany, 1.05438.0500) and incubated at 35°C for 48 h. Fresh subcultures were grown on SDA agar plates and incubated at 35°C for 24 h. Isolated colonies of different pathogenic yeast species grown in SDA medium for 24 h were used to prepare wet mount slides (Fisherbrand, Pittsburg, USA, 22-034486) using sterile distilled water, overlayed with 22 mm coverslips (Fisherbrand, Pittsburg, USA). The raw images for each species were captured using the EVOS M7000 imaging system (Invitrogen, Thermo Fisher Scientific, Massachusetts, USA, AMF7000) with an 100x oil immersion objective l. Coarse and fine adjustments were made to focus the cell on a single plane prior to capturing images. The brightness of the trans-illumination was adjusted to 0.033 before image capture. Images of individual cells were cropped from the raw microscope images to build training, validation, and test datasets 33 . This process was automated using a CNN-based yeast cell segmentation model (yeaZ) for brightfield images 37 . Applying yeaZ produced segmentation masks for each raw microscope image, which contained the boundary coordinates of the identified yeast cells. To ensure each cell was fully contained within its segmentation border, the boundary coordinates were expanded by a factor of 0.15 using morphological dilation (implemented with scikit-image 52 ). The expanded boundary coordinates were used to calculate square (1:1 aspect ratio) bounding boxes around each identified cell, which were then used to crop individual cell images with a padding of 10 pixels. For each crop, pixels that were outside of the expanded cell segmentation boundary coordinates were set to 0 (black) to remove the background around each cell, reducing irrelevant image features. Due to intrinsic errors of the segmentation model and image data, a large number of low-quality images (e.g. slide artifacts and poorly segmented cells) were cropped using this process. To automate the removal of these images, crops that were smaller than 50×50 pixels were excluded (as these were generally found to be slide artifacts), and feature extraction with ResNet18 53 and k-means clustering 54 with k=2 were used to separate the remaining high-quality crops from low-quality crops 55 , which we found to produce consistent results for our datasets. This process was applied to the raw microscope images collected for training, validation, and testing. To ensure dataset balance and avoid biases during the training and validation process 56 , the training and validation sets were randomly downsampled to match the species with the lowest number of crops in each set (each species had a different number of individual cell images due to the random distribution of cells across the raw microscopy images). The test set was randomly downsampled to match the size of the validation set to maintain consistency when comparing model validation and test results. 2.2 Image Dataset Collection A total of 900-1,200 raw microscope images for model training, 100-300 for validation, and 100-250 for testing were collected for each species from different subcultures. The data processing system described in Section 2.1 was applied to the raw images, resulting in 32,809 training, 4,214 validation, and 4,214 test images after k-means to remove low-quality crops and randomly downsampling to maintain balanced datasets. Although k-means removed most low-quality crops, some still appeared in the final datasets, particularly for C. auris and P. kudriavzeviii . This was likely due to the smaller cell size and other morphological differences compared to the yeast species that the segmentation model was trained on ( S. cerevisiae ) 37 , which may have caused challenges during segmentation and clustering. The outlier rates in the test set were quantified by visually inspecting the images and counting the number of outliers for each species, showing that 2.4% of the overall test set consisted of low-quality crops that slipped through detection (705 out of 7 × 4, 214 = 29, 498 images). Most species had an outlier rate of 1% or less, while C. auris and P. kudriavzevii had relatively high rates of 9.1% (387 outliers out of 4,214 images) and 4.8% (204 outliers out of 4,214 images), respectively. To ensure that these outliers (artifacts and poor crops) did not bias our test results, we manually removed them from the test datasets. Model test performance without manual outlier removal was comparable to test performance with manual outlier removal (Supplementary Section 3, Supplementary Figure 5, and Supplementary Table 4). The number of images for each species was then randomly down-sampled to maintain a balanced test set, resulting in 3,827 individual cell images per species. The training, validation, and test raw microscopy images and individual cell crops after each data processing step for each species are summarized in Supplementary Table 1. Pixel and image size distributions for the final training, validation, and test datasets are shown in Supplementary Figures 1, 2, and 3, respectively. 2.3 Computer Vision Model Training and Evaluation Transfer learning and fine-tuning were used to train the computer vision models 57 to classify pathogenic yeasts. The pre-trained base network acts as a generalized model of image classification tasks, while the added network learns features specific to the new training data to enable pathogen classification. To adapt the models to microscopy images of pathogenic yeast species, fine-tuning was applied to the full networks (the base model architecture with a randomly initialized 7-class classification head) by initializing the base model with ImageNetV1 weights and then training the entire network using the microscopy image training dataset. This process adjusted the pre-trained weights to enable the network to learn the visual characteristics specific to the pathogenic yeast species, leveraging the advanced network architectures of the base models and their pre-trained weights to improve the classification performance of the models. Four computer vision architectures were considered for the base model, implemented using PyTorch 20 : InceptionV3 43 , DenseNet121 44 , a base 16 Vision Transformer (ViT) 45 , and a Swin Transformer-Tiny 46 . Image transformations (resizing, rescaling, and standardization) specific to each pre-trained model (based on the ImageNetV1 dataset 58 ) were applied to all images before they were used for training, validation, and testing. Each model was trained on the training set using an AdamW optimizer 59 to minimize a cross-entropy loss function 60 . To ensure consistent model comparisons, the same hyperparameters were used to train each model, with a learning rate of 3e-5, an AdamW weight decay of 0.1, and a batch size of 32. The validation set was used to evaluate model performance (overall accuracy and loss) after each iteration (epoch) of the training process. Early stopping 61 (with a patience of 5) was used to stop the training process once the validation loss no longer decreased, with the model weights from the epoch with the best validation performance (lowest validation loss) saved. After the training and validation process was completed, the models were applied to the test set to evaluate their unbiased generalization performance. Standard classification metrics were used to quantify model performance, including the overall model accuracy and class-specific precision, recall, and F 1 -score, and their averages over all classes. Precision is defined as the ratio between the true positive (TP) and all positive predictions (TP plus false positives) for a given class. Recall and class accuracy are used interchangeably in this study and defined as the correct predictions (TP) divided by the total number of samples (TP plus false negatives) for a given class. The F 1 -score, defined as the harmonic mean between precision and recall, was used to evaluate the balance between these metrics. Confusion matrices were used to visualize model performance and identify classification trends. The classification metrics, confusion matrices, and principal component analysis (PCA) were determined using scikit-learn 62 . To investigate the feature space of the images for each species, we performed a PCA 63 on the test set images after using our trained model as a feature extractor. PCA reduced the dimensionality of the feature representations learned by our trained model to visualize their principal 2D and 3D components. This enabled us to identify differences and similarities between the representations of each species and to better understand common misidentification patterns that occurred ruing model evaluation by showing species that overlap in the feature space. 2.4 Explainable AI for Model Verification and Prediction Interpretation We applied XAI methods to determine if the high-performance computer vision models were learning relevant features (e.g., cellular characteristics such as cell wall and organelles) or irrelevant features (e.g., background artifacts). Additionally, we used XAI to provide additional verification to model predictions and identify features that were most important for species predictions. We chose to use Grad-CAM 50 and Occlusion Sensitivity 51 for our image classification models. Grad-CAM was chosen because it has been extensively tested 64 and used for computer vision model interpretation 65 . Occlusion Sensitivity was selected because it does not rely on specific model architectures 51 and is capable of producing more fine-grained feature importance visualizations than Grad-CAM. Grad-CAM and Occlusion Sensitivity, implemented using the PyTorch-based library Captum 66 , produce normalized attribution maps (values ranging from 0 to 1) that highlight regions of the input image with high importance for a given model prediction. Different sets of test predictions (correct and incorrect) made by the model were investigated by generating Grad-CAM and Occlusion Sensitivity attribution maps for a large number of input images and visually inspecting them to identify image features with the highest attribution (equal to 1). The results of the visual inspection were then consolidated by calculating the proportion of images in the set that contained high attribution on each of the features. First, we used Grad-CAM to distinguish whether there was high attribution inside the cell, outside the cell, or partly inside and outside the cell in each image. This resulted in three measurable binary variables (high attribution yes/1 or no/0 for each case) for each attribution map. Next, Occlusion Sensitivity (with a small non-overlapping 2×2 pixel sliding window) was used to produce fine-grained attribution maps to identify specific image features contributing to model predictions. We identified nine visually distinguishable high-level features that could be present in the input images: internal organelles (nucleus and/or vacuoles), the cell interior (not on a specific organelle), the inner optical pattern, the cell wall, the outer optical pattern, budding patterns/scars, partially cropped cells outside of the main cell, the slide background (between the cell and the background removal border), and the black background. This resulted in nine measurable binary variables for each Occlusion Sensitivity attribution map (high attribution yes/1 or no/0 for each feature), with six of the features being biologically relevant (organelle, cell interior, cell wall, budding pattern/scar, inner and outer optical patterns) and three of these features being irrelevant (other cells in the image, slide background, and black background). As more than one feature had high attribution in the majority of cases, each attribution map generally had more than one high-importance feature. The data collection/processing, model training and evaluation, and XAI workflow are shown in Figure 1 . Download figure Open in new tab Figure 1. Workflow for training and evaluating the advanced computer vision pathogenic yeast species classification models. (a) Raw microscope images for each pathogenic species were collected in three independent groups to create training, validation, and test datasets. Automated data processing was used to crop images of individual cells from the raw microscopy images. (b) The training data set was used to train “black-box” computer vision models, with model performance being evaluated on the validation set after each epoch to optimize hyperparameters. (c) The test data set was used to evaluate model generalization performance, visualized using confusion matrices. Explainable AI was used to interpret model predictions and to identify biologically relevant and irrelevant image features with high prediction importance. Parts of this figure were created using Bio-Render.com. 3 Results and Discussion 3.1 Advanced Computer Vision Pathogenic Yeast Classification We developed high-performance computer vision models to classify seven pathogenic yeast species (Section 2.3). Each computer vision model produced comparable overall test results, with DenseNet121 performing marginally better than the other models ( Figure 2 ). Optimal performance (lowest validation loss) for the DenseNet121 model was obtained after epoch 5 with a training loss, training accuracy, validation loss, and validation accuracy of 0.148, 95%, 0.367, and 88%, respectively ( Figure 3a ). The corresponding validation confusion matrix ( Figure 3b ) and the validation performance metrics (Supplementary Table 2a) were also determined. Applying DenseNet121 to the test set resulted in an overall test accuracy of 87% with strong classification performance across all species, correctly classifying C. albicans, C. auris, N. glabratus, C. haemulonii, P. kudriavzevii, C. parapsilosis , and C. tropicalis in 89%, 86%, 85%, 76%, 89%, 97%, and 82% of cases, respectively ( Figure 4 and Supplementary Table 2b). The test performance was generally balanced between precision and recall across the species, with F 1 scores ranging from 0.83 to 0.92 with an average value of 0.87 (Supplementary Table 2b). The most substantial differences between precision and recall were for C. parapsilosis (0.73 precision, 0.97 recall) and C. haemulonii (0.90 precision, 0.76 recall), with specific predictions contributing to these imbalances ( Figure 4 ). The relatively low precision of C. parapsilosis arose due to other species, primarily N. glabratus, C. haemulonii , and C. auris , being misidentified as C. parapsilosis , while the low recall of C. haemulonii was due to it being misidentified as other species, primarily C. parapsilosis, C. albicans , and N. glabratus . The test set confusion matrices and performance metrics for the other three computer vision models were also obtained (Supplementary Section 2, Supplementary Figure 4, and Supplementary Table 3). All models produced similar test results and classification patterns ( Figure 4 and Supplementary Figure 4), with DenseNet121 producing the best results for N. glabratus and C. haemulonii ( Figure 2 ). To investigate how outliers affect these results when using a fully automated test set without manual outlier removal, we applied the DenseNet121 model to a test set containing outliers (Supplementary Section 3, Supplementary Figure 5, and Supplementary Table 4). The inclusion of outliers resulted in a 1% decrease in overall test accuracy, with a minor effect on performance for all species except C. auris and P. krudiavzevii , which respectively resulted in a 4% decrease in recall and a 4% decrease in precision. Download figure Open in new tab Figure 2. Pathogenic yeast species classification accuracies for advanced computer vision models. DenseNet121, Swin Transformer (Tiny), Vision Transformer (Base 16), and InceptionV3 models had overall test accuracies of 87%, 86%, 86%, and 86%, respectively. Download figure Open in new tab Figure 3. Training and validation performance of the DenseNet121 pathogen classification model. (a) Training and validation accuracy and loss curves. The model weights from the optimal epoch (epoch 5) were saved, which had a validation and training accuracy of 95% and 88% and loss of 0.148 and 0.367, respectively. (b) Confusion matrix for the DenseNet121 model applied to the validation set. The heatmap shows the percentage of samples in each cell relative to the validation dataset size of each class (4214 images). Download figure Open in new tab Figure 4. Confusion matrix of test set predictions made by the DenseNet121 pathogen classification model. The heatmap shows the percentage of samples in each cell relative to the test dataset size of each class (3827 images). To investigate the feature characteristics of the test set for each species used by our DenseNet121 model to make predictions, and to identify reasons for model confusion, we performed a PCA. The PCA revealed that species that were confused by the DenseNet121 model have feature spaces with similar shapes and extensive overlap, for example C. albicans and C. tropicalis ( Figure 5 and Supplementary Video 1). Furthermore, species that the model could distinguish between had minimal overlap in the PCA, including C. albicans and C. auris ( Figure 5 ). In some cases, the model confused one species for another without the reverse being true. For instance N. glabratus was often misidentified as C. parapsilosis , but not vice versa ( Figure 4 . The PCA results indicate that these scenarios occur when a species with a spread-out feature space, such as N. glabratus , overlaps with a species that has a more concentrated feature space, such as C. parapsilosis ( Figure 5 ). This suggests that our DenseNet121 model is biased toward species with more concentrated feature spaces because a lower variance in visible features is easier for the model to learn. This likely occurs due to model confusion when an image of a species with more variability in visible features contains features similar to another species, resulting in inaccurate species predictions. Download figure Open in new tab Figure 5. A 2-dimensional principal component analysis (PCA) of test set images for each pathogenic yeast species using the trained DenseNet121 model as a feature extractor. The first row shows the PCA for individual species, followed by rows showing comparisons of species pairs. A 3-dimensional PCA is provided in Supplementary Video 1. Next, we trained, validated, and tested the high-performance computer vision models using our new approach on four yeast species previously considered to compare with an existing machine learning-based pathogenic yeast classifier 33 . Performance was similar across each of our model architectures, with the Vision Transformer producing the best test results, having an overall accuracy of 92% and correctly classifying C. albicans, C. auris, N. glabratus , and C. haemulonii in 95%, 97%, 96%, and 82% of cases, respectively (Supplementary Section 4 and Supplementary Figure 6). This is a substantial improvement over the previously published test results of the older model, which had an overall accuracy of 78% and correctly classified C. albicans, C. auris, N. glabratus , and C. haemulonii in 97%, 73%, 69%, and 73% of cases, respectively, highlighting the improvements we have made in data collection, processing, and model development. The model architecture test set comparisons, the Vision Transformer confusion matrix, and a species accuracy comparison to previous work 33 are shown in Supplementary Figure 6. Download figure Open in new tab Figure 6. Representative cell images with overlayed Grad-CAM 50 and Occlusion Sensitivity 51 attribution maps for true positive test set predictions made by the DenseNet121 pathogen classification model. The colour bar shows the normalized attribution values. 3.2 Explainable AI to Interpret Computer Vision Model Predictions The size and budding patterns of pathogenic yeast cells were previously hypothesized to contribute correct species classification by computer vision models 33 . However, this hypothesis could not be verified due to the black-box nature of AI models 47 . Here, we used XAI, namely Grad-CAM 50 and Occlusion Sensitivity 51 , to interpret the predictions of high-performance computer vision models. Specifically, we used these XAI methods to identify image features that these improved computer vision models used to accurately identify the species of pathogenic yeast. Specifically, we used Grad-CAM 50 and Occlusion Sensitivity 51 to generate attribution maps for three sets of one hundred randomly selected TP test set predictions for each species made by our best-performing DenseNet121 model. Additionally, we generated attribution maps for the eight most frequent DenseNet121 misidentification cases ( Figure 4 ), namely N. glabratus misidentified as C. parapsilosis, C. haemulonii as C. parapsilosis, C. tropicalis as P. kudriavzevii, C. albicans as C. tropicalis, C. auris as C. parapsilosis, C. haemulonii as C. albicans, C. haemulonii as N. glabratus , and C. tropicalis as C. albicans , which contributed to 12%, 11%, 8%, 6%, 6%, 6%, 4%, and 4% of the test predictions for the corresponding true species, respectively. Representative Grad-CAM and Occlusion Sensitivity attribution maps are shown in Figure 6 , which demonstrate the level of detail and features that each of these methods identify. We calculated the proportion of TP predictions made by DenseNet121 with high Grad-CAM attribution inside the cell, outside of the cell, or both for each pathogenic yeast species ( Figure 7a ). These results showed that there was high prediction importance fully inside the cell in nearly all cases for each species, with no attribution maps indicating high importance fully outside of the cell. For the top eight misidentification cases, the proportion of images with high Grad-CAM attribution was 100% for all species ( Figure 7b ). These results indicate that the model places high prediction importance on the cell in each image in all correct and incorrect predictions investigated. However, the low resolution of the Grad-CAM attribution maps did not allow specific cell characteristics that contribute to model predictions to be identified. Therefore, we applied Occlusion Sensitivity to produce fine-grained attribution maps to identify specific high-importance image features (Section 2.4). Download figure Open in new tab Figure 7. Proportions of identified features with high explainable AI attribution for test set predictions made by the DenseNet121 pathogen classification model. Grad-CAM 50 results are shown for true positive (TP) predictions in (a) and for misidentification predictions in (b) . Occlusion Sensitivity 51 results are shown for TP predictions in (c) and for misidentification predictions in (d) . TP results were averaged over 3 sets of 100 randomly sampled predictions from the test set for each pathogenic yeast species, with the error bars showing the standard deviation. No error bars are shown in the misidentification results, as all images for these cases were analyzed. The proportions of each case do not sum to 100%, as many images had high attribution on more than one feature. Next, we calculated the proportion of TP predictions with high Occlusion Sensitivity attribution on each high-level feature for all species ( Figure 7c ). These results indicate that the DenseNet121model places high importance on biologically relevant features in the majority of TP predictions, while placing little importance on non-biologically relevant features. The most frequent high-importance feature was found to be the cell wall for TP predictions of C. albicans and C. tropicalis , the cell interior (not specific to an organelle) for C. auris, C. haemulonii, P. kudriavzevii , and C. parapsilosis , and internal organelles (nuclei and vacuoles) for N. glabratus . Although these were the features with the most frequent high attribution, the predictions generally had more than one feature with high attribution, and a combination of feature characteristics likely contributed to each model prediction. As for model misidentifications, we again found that the features with high Occlusion Sensitivity attribution were biologically relevant in the vast majority of cases ( Figure 7d ). The cell wall was the most frequent high-importance feature for P. kudriavezevii misidentified as C. tropicalis, C. albicans as C. tropicalis , and C. haemulonii as C. albicans , while the cell interior was the most frequent for the other five misidentification cases. The Grad-CAM and Occlusion Sensitivity XAI methods showed that the cell in each image had the highest importance for species classification predictions made by the DenseNet121 model, indicating that the model relied primarily on biologically relevant cell features to make predictions ( Figures 7 ). The detailed Occlusion Sensitivity attribution maps identified specific biologically relevant cell characteristics and irrelevant image features that had the highest importance for given DenseNet121 predictions ( Figure 7c and 7d), indicating that the interior cell structure, cell wall, and organelles were of the highest importance for distinguishing between the species, followed by the optical patterns around the cells produced by diffraction of the microscope light source and the optical properties of the cells 67 . 4 Conclusion Driven by advances in data collection and processing, model architectures, and training methods, we developed high-performance computer vision models with state-of-the-art generalization performance for human fungal pathogens. Relying only on bright-field microscopy images, this approach does not require sophisticated instruments or expertise and can be integrated into current diagnostic processes 7 . The increase in performance of our advanced computer vision models to more accurately classify a wider range of pathogenic yeasts compared to previous work 31 – 33 was achieved through technical improvements in microscopy image collection, data processing, and model development. Specifically, we obtained a substantially larger training dataset of individual cell images (32,809 per species) with lower time and resource costs compared to previous studies by implementing the yeaZ yeast segmentation model 37 and the k-means algorithm 54,55 to automate raw microscope image processing. Along with full model fine-tuning, this increase in training data enabled our advanced vision models to learn nuanced morphological differences between pathogenic yeast species. The application of XAI methods revealed that biologically relevant features are of high importance in correct and incorrect species predictions made by our advanced computer vision models. Grad-CAM 50 was used for high-level attribution maps, whereas Occlusion Sensitivity 51 was used to elucidate finer details. Grad-CAM showed that computer vision models place high importance on the cell in each image to make predictions, while Occlusion Sensitivity identified important features unique to each pathogenic yeast species, including the cell wall, internal organelles, cell interior structure, and optical properties. The predictions made by our best-performing model, DenseNet121, were determined by the XAI methods to be primarily based on cell morphology, enhancing transparency and trust in model predictions 47 . The XAI methods used in our study generated insights into species-specific features (e.g., optical patterns generated by cellular organelles) that are most important to distinguish yeast species and identified reasons (e.g., morphological similarities between species) for incorrect model predictions. The similarity in performance between the advanced computer vision models developed in our study together with the XAI model interpretation suggests that the high-performance computer vision models are reaching an upper bound of species classification performance on existing microscopy image datasets. Thus, increasing the amount of training data is expected to produce diminishing returns, with further improvements to classification performance requiring advances in experimental data collection and computational data processing to enhance species-distinguishing features, especially when expanding the models to multi-species classification tasks, as observed in other image-based deep learning studies 68 . This may be achieved using high-resolution single-cell imaging, staining methods, or computational feature enhancement. Additional XAI methods could interpret predictions on mixed-species cultures, such as Grad-CAM++ 69 , which has better object localization and explanation of occurrences of multiple objects in a class compared to Grad-CAM 50 . Each pathogenic yeast species considered in this study has a distinct size, shape, cell wall thickness, curvature, internal structure/organelle distribution, budding pattern, and optical properties 70 . This unique distribution of visual features that can be captured through bright-field microscopy, which our advanced computer vision models learned to make accurate species classification predictions. However, there are overlaps in the visual feature distributions ( Figure 5 , Supplementary Video 1, and Figure 7 ), which led to incorrect model predictions when the cell images do not contain sufficient distinguishing information. Other methods are capable of generating unique fingerprints of each species for accurate classification, including MALDI TOF MS by generating mass/charge ratio spectra 10 , Raman spectroscopic methods that detect vibrational signatures of molecules within the yeast cells 18 , using unique volatile organic compounds that are produced by each species 13 , and DNA sequencing 8 . However, we demonstrated that our models can successfully classify pathogenic yeast species based on their unique visual fingerprints captured by bright-field microscopy, while having a substantially lower resource cost compared to these other methods. Our study highlights the power of advanced computer vision models to identify human fungal pathogens and XAI to elucidate the biologically and non-biologically relevant features responsible for model predictions. We anticipate that this approach of combining computer vision models with XAI will advance antimicrobial resistance research and lead to novel diagnostic methods to improve the treatment of patients with infectious diseases. 6. Author Contributions Statement DAC, SAS, and JDG conceptualized the study. JDG carried out the computational work, data visualization, and analytics. SAS carried out the experimental work. SAS and JDG interpreted the results. JDG, SAS, and DAC wrote the manuscript. DAC acquired the funding and supervised the research. 7 Data Availability The data that support the findings of this study are available upon reasonable, noncommercial request from the corresponding author. 8 Competing Interests D.A.C. is funded by an AICE-Concepts grant to develop a commercial point-of-care medical diagnostic device based on artificial intelligence. 5 Acknowledgments This research was funded by grants to D.A.C. from the Human Frontier Science Program (RGEC30/2024) and the Alberta Innovates Accelerating Innovations into Care - Concepts program (AICE-Concepts 597389). We thank Dr. Tanis Dingle and the Alberta Precision Laboratories-Public Health Laboratory for providing the Candida species isolates, Asst. Prof. Michael Manhart for suggesting the use of principal component analysis, and Prof. Randy Goebel for helpful discussions on explainable AI. Funder Information Declared International Human Frontier Science Program Organization, https://ror.org/02ebx7v45 , RGEC30/2024 Alberta Innovates , AICE-Concepts 597389 References 1. ↵ Denning , D. W. Global incidence and mortality of severe fungal disease . The Lancet Infect. Dis . 24 , e428 – e438 ( 2024 ). OpenUrl CrossRef 2. ↵ Corrêa-Junior , D. , Andrade , I. B. d. , Alves , V. , Araújo , G.R.d.S. & Frases , S. Clinical challenges of emerging and re-emerging yeast infections in the context of the covid-19 pandemic . Microorganisms 10 , 2223 ( 2022 ). OpenUrl CrossRef PubMed 3. ↵ Pappas , P. G. , Lionakis , M. S. , Arendrup , M. C. , Ostrosky-Zeichner , L. & Kullberg , B. J. Invasive candidiasis . Nat. Rev. Dis. Primers 4 , 1 – 20 ( 2018 ). OpenUrl PubMed 4. ↵ Cornely , O. A. et al. Global guideline for the diagnosis and management of candidiasis: an initiative of the ecmm in cooperation with isham and asm . The Lancet Infect. Dis . ( 2025 ). 5. ↵ CDC . Data and statistics on candidemia . https://www.cdc.gov/candidiasis/data-research/facts-stats/index.html ( 2024 ). Accessed: (6 April 2025 ). 6. ↵ CDC . Increasing threat of spread of antimicrobial-resistant fungus in healthcare facilities . https://www.cdc.gov/media/releases/2023/p0320-cauris.html ( 2023 ). Accessed: (6 April 2025 ). 7. ↵ Mendonca , A. , Santos , H. , Franco-Duarte , R. & Sampaio , P. Fungal infections diagnosis–past, present and future . Res. microbiology 173 , 103915 ( 2022 ). OpenUrl CrossRef 8. ↵ Iyer , V. et al. Culture-independent identification of bloodstream infections from whole blood: Prospective evaluation in specimens of known infection status . J. Clin. Microbiol . 62 , e01498 – 23 ( 2024 ). OpenUrl PubMed 9. ↵ Birinci , A. , Akkurt , L. , Acuner , C. , Unlu , M. & Durupinar , B. Rapid identification of the candida species from direct blood cultures by chromagar™ candida . J. international medical research 32 , 484 – 487 ( 2004 ). OpenUrl CrossRef 10. ↵ Ashutosh Panda , A. P. et al. Maldi-tof mass spectrometry for rapid identification of clinical fungal isolates based on ribosomal protein biomarkers . ( 2015 ). 11. ↵ Gudisa , R. & Rudramurthy , S. M. Role of gene sequencing for the diagnosis, tracking and prevention of fungal infections . J. The Acad. Clin. Microbiol . 24 , S15 – S24 ( 2022 ). OpenUrl CrossRef 12. ↵ Wilson , N. M. et al. T2 magnetic resonance assay improves timely management of candidemia . JAMS 1 , 12 – 18 ( 2017 ). OpenUrl 13. ↵ Bastos , M. L. et al. Breaking barriers in candida spp. detection with electronic noses and artificial intelligence . Sci. Reports 14 , 956 ( 2024 ). OpenUrl 14. ↵ Franconi , I. et al. Rapid identification of clinically relevant candida spp. by i-done software using attenuated total reflectance fourier transform infrared (atr-ftir) spectroscopy . J. Fungi 11 , 40 ( 2025 ). OpenUrl CrossRef 15. ↵ Higashi , Y. , Niimi , H. , Sakamaki , I. , Yamamoto , Y. & Kitajima , I. Rapid identification of candida species in candidemia directly from blood samples using imperfect match probes . Sci. Reports 10 , 5828 ( 2020 ). OpenUrl 16. ↵ Wang , L. et al. Rapid detection of candida tropicalis in clinical samples from different sources using rpa-lfs . Front. Cell. Infect. Microbiol . 12 , 898186 ( 2022 ). OpenUrl CrossRef PubMed 17. ↵ Nemcova , E. et al. Rapid identification of medically important candida isolates using high resolution melting analysis . PLoS One 10 , e0116940 ( 2015 ). OpenUrl CrossRef PubMed 18. ↵ Koya , S. K. et al. Rapid identification of candida auris by raman spectroscopy combined with deep learning . J. Raman Spectrosc . 56 , 218 – 227 ( 2025 ). OpenUrl CrossRef 19. ↵ Rehse , S. J. A review of the use of laser-induced breakdown spectroscopy for bacterial classification, quantification, and identification . Spectrochimica Acta Part B: At. Spectrosc . 154 , 50 – 69 ( 2019 ). OpenUrl CrossRef 20. ↵ Paszke , A. et al. Pytorch: An imperative style, high-performance deep learning library . Adv. neural information processing systems 32 ( 2019 ). 21. ↵ Xu , Y. et al. Artificial intelligence: A powerful paradigm for scientific research . The Innov . 2 ( 2021 ). 22. ↵ Abramson , J. et al. Accurate structure prediction of biomolecular interactions with alphafold 3 . Nature 630 , 493 – 500 ( 2024 ). OpenUrl CrossRef PubMed 23. ↵ Zhang , X. , Zhang , D. , Zhang , X. & Zhang , X. Artificial intelligence applications in the diagnosis and treatment of bacterial infections . Front. Microbiol . 15 , 1449844 ( 2024 ). OpenUrl CrossRef PubMed 24. ↵ Zheng , E. J. et al. Discovery of antibiotics that selectively kill metabolically dormant bacteria . Cell Chem. Biol . 31 , 712 – 728 ( 2024 ). OpenUrl CrossRef PubMed 25. ↵ Guo , M.-H. et al. Attention mechanisms in computer vision: A survey . Comput. visual media 8 , 331 – 368 ( 2022 ). OpenUrl CrossRef 26. ↵ Byerly , A. , Kalganova , T. & Ott , R. The current state of the art in deep learning for image classification: A review . In Science and information conference , 88 – 105 ( Springer , 2022 ). 27. ↵ Zhao , X. et al. A review of convolutional neural networks in computer vision . Artif. Intell. Rev . 57 , 99 ( 2024 ). OpenUrl CrossRef 28. ↵ Chai , J. , Zeng , H. , Li , A. & Ngai , E. W. Deep learning in computer vision: A critical review of emerging techniques and application scenarios . Mach. Learn. with Appl . 6 , 100134 ( 2021 ). OpenUrl CrossRef 29. ↵ Ali , M. et al. Applications of artificial intelligence, deep learning, and machine learning to support the analysis of microscopic images of cells and tissues . J. Imaging 11 , 59 ( 2025 ). OpenUrl CrossRef PubMed 30. ↵ Esteva , A. et al. Deep learning-enabled medical computer vision . NPJ digital medicine 4 , 5 ( 2021 ). OpenUrl CrossRef PubMed 31. ↵ Zieliński , B. , Sroka-Oleksiak , A. , Rymarczyk , D. , Piekarczyk , A. & Brzychczy-Wloch , M. Deep learning approach to describe and classify fungi microscopic images . PloS one 15 , e0234806 ( 2020 ). OpenUrl CrossRef PubMed 32. Rahman , M. A. et al. Classification of fungal genera from microscopic images using artificial intelligence . J. Pathol. Informatics 14 , 100314 ( 2023 ). OpenUrl CrossRef 33. ↵ Shankarnarayan , S. A. & Charlebois , D. A. Machine learning to identify clinically relevant candida yeast species . Med. Mycol . 62 , myad134 ( 2024 ). OpenUrl CrossRef PubMed 34. ↵ Guo , Y. , Liu , Y. , Georgiou , T. & Lew , M. S. A review of semantic segmentation using deep neural networks . Int. journal multimedia information retrieval 7 , 87 – 93 ( 2018 ). OpenUrl CrossRef 35. ↵ Stringer , C. & Pachitariu , M. Cellpose3: One-click image restoration for improved cellular segmentation. biorxiv 2024 . bioRxiv ( 2010 ). 36. Schmidt , U. , Weigert , M. , Broaddus , C. & Myers , G. Cell detection with star-convex polygons . In Medical image computing and computer assisted intervention–MICCAI 2018: 21st international conference, Granada, Spain, September 16-20, 2018, proceedings, part II 11 , 265 – 273 ( Springer , 2018 ). 37. ↵ Dietler , N. et al. A convolutional neural network segments yeast microscopy images with high accuracy . Nat. communications 11 , 5723 ( 2020 ). OpenUrl CrossRef 38. ↵ Bunk , D. et al. Yeastmate: neural network-assisted segmentation of mating and budding events in saccharomyces cerevisiae . Bioinformatics 38 , 2667 – 2669 ( 2022 ). OpenUrl CrossRef PubMed 39. Salem , D. et al. Yeastnet: Deep-learning-enabled accurate segmentation of budding yeast cells in bright-field microscopy . Appl. Sci . 11 , 2692 ( 2021 ). OpenUrl CrossRef 40. ↵ Lee , M. , Kunzi , M. , Neurohr , G. , Lee , S. S. & Park , Y. Hybrid machine-learning framework for volumetric segmentation and quantification of vacuoles in individual yeast cells using holotomography . Biomed. Opt. Express 14 , 4567 – 4578 ( 2023 ). OpenUrl CrossRef PubMed 41. ↵ Szabó , G. , Bonaiuti , P. , Ciliberto , A. & Horváth , A. Enhancing yeast cell tracking with a time-symmetric deep learning approach . npj Syst. Biol. Appl . 11 , 25 ( 2025 ). OpenUrl CrossRef PubMed 42. ↵ Lawson , K. , Skrtic , S. , Vo , M. & Escorcia , W. Measuring cell dimensions in fission yeast using machine learning . In Schizosaccharomyces pombe: Methods and Protocols , 33 – 46 ( Springer , 2024 ). 43. ↵ Szegedy , C. , Vanhoucke , V. , Ioffe , S. , Shlens , J. & Wojna , Z. Rethinking the inception architecture for computer vision . In Proceedings of the IEEE conference on computer vision and pattern recognition , 2818 – 2826 ( 2016 ). 44. ↵ Huang , G. , Liu , Z. , Van Der Maaten , L. & Weinberger , K. Q. Densely connected convolutional networks . In Proceedings of the IEEE conference on computer vision and pattern recognition , 4700 – 4708 ( 2017 ). 45. ↵ Dosovitskiy , A. An image is worth 16×16 words: Transformers for image recognition at scale . arXiv preprint arxiv: 2010.11929 ( 2020 ). 46. ↵ Liu , Z. et al. Swin transformer: Hierarchical vision transformer using shifted windows . In Proceedings of the IEEE/CVF international conference on computer vision , 10012 – 10022 ( 2021 ). 47. ↵ Durán , J. M. & Jongsma , K. R. Who is afraid of black box algorithms? on the epistemological and ethical basis of trust in medical ai . J. medical ethics 47 , 329 – 335 ( 2021 ). OpenUrl Abstract / FREE Full Text 48. ↵ Dwivedi , R. et al. Explainable ai (xai): Core ideas, techniques, and solutions . ACM Comput. Surv . 55 , 1 – 33 ( 2023 ). OpenUrl CrossRef 49. ↵ Kamakshi , V. & Krishnan , N. C. Explainable image classification: The journey so far and the road ahead . AI 4 , 620 – 651 ( 2023 ). OpenUrl CrossRef 50. ↵ Selvaraju , R. R. et al. Grad-cam: Visual explanations from deep networks via gradient-based localization . In Proceedings of the IEEE international conference on computer vision , 618 – 626 ( 2017 ). 51. ↵ Zeiler , M. Visualizing and understanding convolutional networks . In European conference on computer vision/arXiv , vol. 1311 ( 2014 ). 52. ↵ Van der Walt , S. et al. scikit-image: image processing in python . PeerJ 2 , e453 ( 2014 ). OpenUrl CrossRef PubMed 53. ↵ He , K. , Zhang , X. , Ren , S. & Sun , J. Deep residual learning for image recognition . In Proceedings of the IEEE conference on computer vision and pattern recognition , 770 – 778 ( 2016 ). 54. ↵ Steinley , D. K-means clustering: a half-century synthesis . Br. J. Math. Stat. Psychol . 59 , 1 – 34 ( 2006 ). OpenUrl CrossRef PubMed Web of Science 55. ↵ Guérin , J. , Thiery , S. , Nyiri , E. , Gibaru , O. & Boots , B. Combining pretrained cnn feature extractors to enhance clustering of complex natural images . Neurocomputing 423 , 551 – 571 ( 2021 ). OpenUrl CrossRef 56. ↵ Kumar , P. , Bhatnagar , R. , Gaur , K. & Bhatnagar , A. Classification of imbalanced data: review of methods and applications . In IOP conference series: materials science and engineering , vol. 1099 , 012077 ( IOP Publishing , 2021 ). OpenUrl CrossRef 57. ↵ Shaha , M. & Pawar , M. Transfer learning for image classification . In 2018 second international conference on electronics, communication and aerospace technology (ICECA) , 656 – 660 ( IEEE , 2018 ). 58. ↵ Deng , J. et al. Imagenet: A large-scale hierarchical image database . In 2009 IEEE conference on computer vision and pattern recognition , 248 – 255 ( Ieee , 2009 ). 59. ↵ Loshchilov , I. Decoupled weight decay regularization . arXiv preprint arxiv: 1711.05101 ( 2017 ). 60. ↵ Mao , A. , Mohri , M. & Zhong , Y. Cross-entropy loss functions: Theoretical analysis and applications . In International conference on Machine learning , 23803 – 23828 ( PMLR , 2023 ). 61. ↵ Prechelt , L. Early stopping-but when? In Neural Networks: Tricks of the trade , 55 – 69 ( Springer , 2002 ). 62. ↵ Pedregosa , F. et al. Scikit-learn: Machine learning in python . J. machine Learn. research 12 , 2825 – 2830 ( 2011 ). OpenUrl 63. ↵ Greenacre , M. et al. Principal component analysis . Nat. Rev. Methods Primers 2 , 100 ( 2022 ). OpenUrl CrossRef 64. ↵ Adebayo , J. et al. Sanity checks for saliency maps . Adv. neural information processing systems 31 ( 2018 ). 65. ↵ Van der Velden , B. H. , Kuijf , H. J. , Gilhuijs , K. G. & Viergever , M. A. Explainable artificial intelligence (xai) in deep learning-based medical image analysis . Med. Image Analysis 79 , 102470 ( 2022 ). OpenUrl CrossRef 66. ↵ Kokhlikyan , N. et al. Captum: A unified and generic model interpretability library for pytorch . arXiv preprint arxiv: 2009.07896 ( 2020 ). 67. ↵ Yu Wan , W. et al. Integration of light scattering with machine learning for label free cell detection . Biomed. Opt. Express 12 , 3512 – 3529 ( 2021 ). OpenUrl CrossRef PubMed 68. ↵ Bai , B. et al. Deep learning-enabled virtual histological staining of biological samples . Light. Sci. & Appl . 12 , 57 ( 2023 ). OpenUrl CrossRef PubMed 69. ↵ Chattopadhay , A. , Sarkar , A. , Howlader , P. & Balasubramanian , V. N. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks . In 2018 IEEE winter conference on applications of computer vision (WACV) , 839 – 847 ( IEEE , 2018 ). 70. ↵ Walsh , T. J. , Hayden , R. T. & Larone , D. H. Larone’s medically important fungi: A guide to identification , 121 – 146 ( John Wiley & Sons , 2018 ). View the discussion thread. Back to top Previous Next Posted June 29, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Peering Inside the Black Box: Explainable AI to Interpret Advanced Computer Vision Fungal Pathogen Prediction Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Peering Inside the Black Box: Explainable AI to Interpret Advanced Computer Vision Fungal Pathogen Prediction Joshua D. Guthrie , Shamanth A. Shankarnarayan , Daniel A. Charlebois bioRxiv 2025.06.27.662051; doi: https://doi.org/10.1101/2025.06.27.662051 Share This Article: Copy Citation Tools Peering Inside the Black Box: Explainable AI to Interpret Advanced Computer Vision Fungal Pathogen Prediction Joshua D. Guthrie , Shamanth A. Shankarnarayan , Daniel A. Charlebois bioRxiv 2025.06.27.662051; doi: https://doi.org/10.1101/2025.06.27.662051 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Microbiology Subject Areas All Articles Animal Behavior and Cognition (7633) Biochemistry (17680) Bioengineering (13889) Bioinformatics (41928) Biophysics (21445) Cancer Biology (18585) Cell Biology (25491) Clinical Trials (138) Developmental Biology (13373) Ecology (19897) Epidemiology (2067) Evolutionary Biology (24308) Genetics (15606) Genomics (22496) Immunology (17736) Microbiology (40385) Molecular Biology (17175) Neuroscience (88583) Paleontology (666) Pathology (2830) Pharmacology and Toxicology (4822) Physiology (7641) Plant Biology (15149) Scientific Communication and Education (2045) Synthetic Biology (4293) Systems Biology (9822) Zoology (2271)