Fungal infections rank among the most underappreciated threats in the context of global health. In 2022, the World Health Organization (WHO) published its first Fungal Priority Pathogens List, which cataloged 19 fungi across critical, high, and medium priority tiers, signaling the urgent need to strengthen surveillance, prevention, and diagnostic capacity1. The associated disease burden is considerable: millions of infections take place worldwide each year2, and more recent conclusions estimate approximately 3.8 million related deaths annually, of which roughly 2.5 million are directly attributable to invasive fungal diseases3. This toll rivals or exceeds the mortality that is attributed to tuberculosis, and it is several times that of malaria. The total overall mortality is almost certainly underestimated because diagnostic limitations trans- late directly into missed and delayed diagnoses.

Each pillar of conventional fungal diagnosis carries well-recognized constraints. Direct microscopy—potassium hy- droxide (KOH) wet mounts, calcofluor white staining, and histopathology—but exhibits highly variable sensitivity (20-80%) and pronounced operator dependence4. Pathogen culturing, although this continues to be regarded as a reference standard, requires a prolonged period of incubation that can extend from days to weeks, only poorly recovers many fastidious or non-viable organisms, and provides limited sen- sitivity for several specimen types5. Polymerase chain reaction and related molecular assays can improve sensitivity but are typically restricted to predefined targets and remain unevenly available. Such serological biomarkers as galactomannan and (1→3)-β-D-glucan can support diagnosis but lack species-level specificity and perform inconsistently across host populations6. Collectively, these gaps require clinicians to rely on empirical therapy during the very window when timely, targeted treatment most influences outcome.

Artificial intelligence (AI) and machine learning (ML) have emerged as computational strategies that are well suited to addressing these shortcomings. AI systems that learn com- plex, high-dimensional relationships from large datasets can automate image interpretation, extract diagnostic signals from mass spectra or genomic sequences, fuse information across modalities, and deliver rapid, reproducible inferences that are less susceptible to interobserver variation7. Over the past five years, this activity has accelerated markedly, with applications now spanning virtually every diagnostic modality relevant to clinical mycology8.

This narrative review surveys the recent evolution of AI-based diagnostic approaches for fungal infections, considering their prospects for clinical translation. Instead of providing an exhaustive, systematic enumeration, we adopt a thematic and interpretive approach that draws on representative studies published between 2020 and 2025 to map the landscape, appraise reported performance critically, and identify the barriers separating proof-of-concept from bedside utility. We organize the discussion into 10 thematic domains: micro- scopic image analysis; CT radiomics for pulmonary myco-ses; ophthalmic mycology; dermatological mycology; mass spectrometry; vibrational (Raman) spectroscopy; biosensor-based detection of volatile organic compounds (VOCs); meta- genomic next-generation sequencing (mNGS) and AI inte- gration; prediction of antifungal susceptibility; and integration of multimodal data. For each, we examine the AI architectures that were employed, their reported diagnostic performance, the nature of the underlying datasets, and the challenges to validation that remain, before turning to the cross-cutting obstacles that must be overcome for these technologies to enter routine practice.

AI-based diagnostic systems for fungal infection generally draw on a broad repertoire of contemporary computational methods. Convolutional neural networks (CNNs)—including such widely used architectures as VGG, ResNet, DenseNet, Inception, and EfficientNet—have shown particular promise in image-based tasks due to their capacity to learn hierarchical features directly from visual data7-9. Object-detection frame- works, notably the You Only Look Once family and Faster Region-based CNN (Faster R-CNN), have been used to detect and localize fungal structures in microscopic and dermoscopic images10,11. Because annotated mycological datasets are scarce, transfer learning—fine-tuning models that are pre-trained on large general-purpose image collections such as ImageNet—has become a common and pragmatic solution8. Classical ML retains an important role, especially when inputs are structured, such as in the context of radiomic features, spectral data, or clinical variables. Such methods as Random Forest, support vector machines, gradient-boosting algo- rithms, and logistic regression have been applied within these settings7,8. Ensemble strategies combining multiple classifiers frequently prove more robust than any single model. More recently, attention mechanisms, vision transformers, and foundation models that are pre-trained on large biomedical images are beginning to be explored for mycological appli- cations, although this work continues to be at an early stage7,8.

A useful conceptual distinction separates fully supervised, end-to-end deep-learning systems from hybrid pipelines where hand-crafted features (for example, radiomics) are extracted and then classified using classical ML. Both have achieved competitive performance, and the optimal choice is generally dictated by dataset size, available computational resources, and the premium placed on interpretability7,8.

3.1 Fluorescence and brightfield microscopy

Microscopy of clinical specimens continues to be a first-line approach to investigating fungal infection, and several recent studies have examined AI-based fungal detection in microscopic images12,13. He et al. developed an artificial-intelligence-powered fluorescence microscopic image analyzer (FMIA) for the automated detection of superficial fungal infections. In a study of 300 patients (241 mycologically confirmed), it achieved a sensitivity of 96.27%, a specificity of 94.92%, and an area under the receiver operating characteristic curve (AUC) of 0.96, outperforming potassium hydroxide microscopy and matching fluorescence staining12. This performance is especially attractive for high-throughput screening within busy clinical microbiology laboratories, where manual review is labor-intensive and subject to interobserver variability, and it demonstrates that real-time object-detection architectures can be successfully adapted to mycology.

The difficulty of fine-grained morphological classification is, however, equally instructive. Rahman et al. (2023) trained CNNs to classify brightfield images that span 89 fungal genera; their best-performing DenseNet model reached a top-1 accuracy of 65.35% and a top-3 accuracy of 75.19%, values that underscore the inherent challenge of large-scale, genus-level discrimination and the gap remaining between binary detection and comprehensive taxonomic identification13.

3.2 Colony morphology and culture-plate analysis

A complementary strategy targets macroscopic colony morphology on culture plates. Tsang et al. (2025) built an image-recognition system for identifying clinically important Aspergillus species where a ResNet-18 model achieved a testing accuracy of 99.35%, suggesting that accurate species-level identification could be obtained without the use of MALDI-TOF MS or molecular methods14. This is a particularly relevant proof-of-concept for resource-limited laboratories lacking access to more advanced identification platforms. Efforts to standardize this work are also underway: the publicly available OpenFungi benchmark dataset (2024) was created to support the reproducible, standardized evaluation of fungal image-recognition models across phylogenetically diverse species15.

Invasive pulmonary mycoses—including invasive pulmonary aspergillosis (IPA), pulmonary mucormycosis, and Pneumo- cystis jirovecii pneumonia (PJP)—frequently produce over- lapping appearances within computed tomography (CT), rendering it difficult to produce confident diagnoses even for experienced radiologists. AI-based analysis of chest CT has consequently attracted attention as a means of improving the early detection and differentiation of fungal lung infection16-18.

4.1 IPA

Zhang et al. (2024) developed a combined clinical-radiomics-deep-learning model for IPA in a retrospective co- hort of 263 patients (148 with IPA, 115 without)16. Radiomic features that were extracted from CT were integrated with clinical risk factors and deep-learning features that were derived from four pre-trained CNNs; the combined model achieved an AUC of 0.881 in the test set, outperforming both clinical-radiological and radiomics-only models16.

Liu et al. (2024) introduced MI-DenseCFNet, a multimodal architecture fusing high-resolution CT (HRCT) features with clinical data to distinguish Aspergillus pneumonia from Staphylococcus aureus pneumonia17. The model produced an AUC of 0.92 on internal validation and 0.83 on external validation; the diagnostic accuracy improved from 78.4% with imaging features alone to 88.1% after the incorporation of clinical data17. Extending this to a more demanding multi-class problem, Li et al. developed deep-learning segmentation models that are capable of differentiating IPA, mucormycosis, bacterial pneumonia, and pulmonary tuberculosis with respect to CT, demonstrating the feasibility of complex differential diagnosis of pulmonary infiltrates in immunocompromised patients18.

4.2 Differentiating pulmonary fungal infections

The distinction of IPA from PJP has direct therapeutic consequences, as the two require fundamentally different management. Peng et al. (2025) created a CT-based radio- mics model for this distinction in 97 cases, examining six supervised classifiers; XGBoost performed best in terms of validation (AUC 0.808; 95% CI 0.655-0.961)19. Notably, classifiers that are trained on regions of interest that en- compass both consolidation and adjacent ground-glass opacity outperformed those that were based on consolidation alone, highlighting the diagnostic importance of peri-lesional features within fungal lung infection19.

Despite these encouraging results, a large majority of AI-based CT radiomics studies have been retrospective and fea- tured only a single center. Heterogeneity in imaging protocols, scanner hardware, and patient populations threatens genera- lizability across institutions, and large multicenter datasets with prospective validation are needed to establish genuine clinical utility.

Fungal keratitis is a potentially blinding corneal infection that is most commonly found in rural and agricultural communities, where delayed diagnosis frequently culminates in irreversible loss of vision. In vivo confocal microscopy (IVCM) is a valuable non-invasive diagnostic tool, and AI is increasingly being used to automate the interpretation of the images it generates.

5.1 Evidence from systematic reviews and meta-analyses

Assaf et al. (2025) reviewed computer-vision applications for infectious keratitis, describing the expanding role played by deep learning in the automated detection and classification of corneal imaging findings, emphasizing the need for stand- ardized datasets, external validation, and multicenter studies to support clinical translation20.

Ong et al. (2024) performed a systematic review and meta-analysis of 35 studies that incorporated 136,401 images from more than 56,011 patients21. AI-based tools demonstrated strong performance, with a pooled sensitivity and specificity of 86.2% and 96.3%, respectively, in external validation. The analysis also exposed a substantial geographic skew that could limit generalizability: roughly 60% of the studies that were included originated in centers in China21.

5.2 Deep learning for IVCM-based fungal keratitis diagnosis

Li et al. (2024) adopted a two-stage deep neural network for fungal keratitis using IVCM images. It was trained and tested on 96,632 images, and it achieved a sensitivity of 97.57% and a specificity of 96.65%, comparable to or ex- ceeding the performance of experienced ophthalmologists22. Recognizing that diagnostic accuracy alone is insufficient for clinical adoption, Essalat et al. (2023) developed an inter- pretable deep-learning approach to differentiate fungal from Acanthamoeba keratitis in IVCM images, incorporating class-activation mapping to render model predictions visually explicable to clinicians23.

5.3 Multimodal methods for keratitis diagnosis

Prajna et al. (2025) reported a prospective study of 599 patients using computer-vision and multimodal machine-learning models to differentiate bacterial and fungal keratitis. The best-performing computer-vision model achieved an AUC of 0.81, an accuracy of 77%, and an F1-score of 0.85, whereas the multimodal model achieved an AUC of 0.82, an accuracy of 81%, and an F1-score of 0.8924. The study is notable as being among the few prospective validations in a field that is still dominated by retrospective work, and the coupling of clinical metadata with imaging is a meaningful step toward deployable decision-support systems.

Nevertheless, the high diagnostic performance reported across ophthalmic AI studies is tempered by the use of per- sistent geographic bias, with most datasets drawn from a small number of countries. Broader international collaboration and prospective clinical validation are essential before these tools can be applied confidently across diverse populations.

Superficial fungal infections—onychomycosis, tinea corporis, and tinea capitis—are among the most common dermato- logical conditions worldwide, which makes them an attractive target for automated screening. Kim et al. (2024) reviewed emerging applications of AI in medical mycology, including the automated analysis of KOH microscopy, fungal culture, histopathology, and clinical images. They also highlighted potential and current limitations of AI-assisted fungal diag- nosis25. A comprehensive review performed by Hasan Pour (2025) that summarizes 54 studies of AI applications for superficial fungal infection provides a broad overview of the field26. Several studies have reported AI performance that exceeds that of clinicians in their evaluated datasets. Zhu et al. (2022) trained Faster R-CNN models using dermoscopic images, including 603 images from patients who have onycho- mycosis and developed a multi-step pipeline to differentiate nail disorders and recognize specific dermoscopic patterns. Their ensemble model produced an accuracy of 87.5%, a sensitivity of 78.5%, and a specificity of 93.0%, outperforming a panel of 54 dermatologists11. Yilmaz et al. (2022) used VGG16 and InceptionV3 on KOH microscopy images, attaining mean accuracies of 88.10% and 88.78%, respectively, against 74.53% for 16 dermatologists27.

Complementing these findings, Cinar and Taspinar (2023) demonstrated the feasibility of CNN-based classification of microscopic fungal images, reinforcing the potential of AI-assisted image analysis for rapid screening in routine laboratory settings9. Koo et al. developed a YOLOv4-based regional CNN for the automated detection of fungal hyphae in KOH microscopy images, demonstrating high sensitivity and specificity across different magnifications10. These com- parisons are encouraging, although the apparent superiority of AI needs to be interpreted cautiously, as it is typically established in curated, single-center datasets rather than in prospective clinical use.

Indeed, most dermatological AI studies are limited by small datasets, single-center designs, and restricted geographic diversity. Variability in image-acquisition conditions, skin pig- mentation, lesion morphology, and the regional distribution of fungal species can affect generalizability, and multicenter validation having larger annotated datasets and prospective evaluation will be required to confirm real-world utility.

The use of AI in dermatological mycology extends to teledermatology and histopathological assessment. Muñoz-López et al. evaluated a deep neural network in a prospective teledermatology study including fungal diagnoses such as tinea and onychomycosis; that study demonstrated the potential of AI-assisted image interpretation in a real-world dermatological consultation setting28. Deep learning has also been investigated for the histopathological diagnosis of onychomycosis. Decroos et al. developed a deep-learning system for the detection of fungal elements in histological nail-clipping specimens and reported non-inferiority to analog diagnosis by histopathologists for specificity and an AUC of 0.98129. Similarly, Jansen et al. evaluated a U-Net-based segmentation model using histological whole-slide images and reported a fungal-element detection sensitivity that was comparable to that of board-certified dermatopathologists, indicating its potential as an assistive screening tool30. These studies broaden the role of AI in dermatological mycology that goes beyond dermoscopic and KOH microscopy image classification while also highlighting the need for prospective, multicenter validation to precede clinical implementation.

7.1 MALDI-TOF mass spectrometry

MALDI-TOF MS has become an important tool for the rapid identification of clinically relevant fungi, including yeasts and filamentous fungi. The technology offers rapid turnaround times, high accuracy, and reduced laboratory costs compared with conventional phenotypic identification methods. However, identification performance remains dependent on database quality and specimen processing protocols31.

AI-assisted spectral analysis has been proposed as a route to extend fungal identification beyond conventional database matching32. Normand et al. (2022) applied CNNs to MALDI-TOF spectra to identify clonal populations of Aspergillus flavus, achieving over 93% accuracy on two of the three mass spectrometry instruments tested, with reduced accuracy on a third, older device, and demonstrating the feasibility of AI-driven spectral analysis for strain-level characterization32.

7.2 Single-cell Raman spectroscopy

Single-cell Raman spectroscopy (SCRS) implemented with AI has recently emerged as a promising platform for the development of rapid, culture-independent fungal identifi- cation33,34. Xu et al. (2023) established an AI-based SCRS pipeline that was validated on 115,129 single-cell spectra that were collected from clinical fungal isolates obtained from 94 patients33. The system identified fungal species with 100% accuracy using five cells per patient, requiring a 2-second acquisition per cell and an overall sample-to-result time of approximately 1 hour—a striking contrast with culture-based identification, which commonly requires 48-72 hours or longer33.

SCRS also shows utility for antifungal susceptibility testing. A 2024 study of Candida auris reported 93.33% accuracy for species-level characterization and prediction accuracies of 99% and 94% for fluconazole and amphotericin B resistance, respectively34. Because C. auris is a globally emerg- ing, multidrug-resistant organism and a critical threat to public health, the ability to derive both identity and susceptibility from a single rapid assay has important implications for management and antimicrobial stewardship. Functioning at the single-cell level, SCRS is inherently culture-independent and can circumvent the roughly 50% rate of obtaining negative cultures that hampers conventional fungal diagnosis33.

Despite its remarkable speed and accuracy, the broad adoption of SCRS is likely to be constrained by the cost of the instrumentation, the technical expertise that it demands, and its limited availability in routine laboratories. Standardized analytical workflows and larger validation studies are pre- requisites for wider clinical deployment.

The profiling of VOCs using electronic-nose technology is increasingly explored for rapid microbial detection and clas- sification35,36. Bastos et al. (2024) coupled an electronic-nose platform with an InceptionTime deep-learning model to identify six Candida species, achieving accuracy, precision, recall, and F1-score values that were all above 95%35. Castro et al. (2022) reported comparable feasibility, combining VOC analysis with ML to identify three Candida species with greater than 90% accuracy36. The low cost and portability of these devices make them attractive for point-of-care applications in particular in resource-constrained settings35,36.

Nevertheless, environmental conditions, sensor drift, and variability in VOC profiles can all influence performance. Further standardization and real-world validation are neces- sary to ensure reproducibility across diverse healthcare settings.

Metagenomic next-generation sequencing (mNGS) enables unbiased, hypothesis-free detection of microbial nucleic acid in clinical specimens and can identify fungi that are missed by culturing methods and targeted molecular assays. Miao et al. (2018) showed that mNGS substantially improved the detection of pathogens in clinical infectious-disease practice and offered broader diagnostic coverage than conventional methods37. While mNGS is not itself an AI technology, ML algorithms are increasingly embedded in its workflow for taxonomic classification, background-noise filtering, and the interpretation of clinically relevant fungal signals.

A central challenge for mNGS-based fungal diagnosis is drawing a distinction between true infection and colonization and environmental contamination. Jiang et al. (2024) analyzed the sequence abundance of Aspergillus in bronchoalveolar lavage fluid and found significant differences between in- fection and colonization (p < 0.0001), with mNGS achieving an AUC of 0.894 (95% CI 0.811-0.976) for this distinction38. Yin et al. (2025) further demonstrated that plasma mNGS improved the diagnosis of fungal infection and supported the optimization of antifungal therapy in immunocompro- mised patients39. Together, these findings suggest that the integration of mNGS with advanced computational analytics is approaching clinical applicability for selected high-risk populations.

The interpretation of such findings remains difficult because detected fungal DNA could reflect infection, colonization, or contamination. High sequencing costs, demanding bio-informatic requirements, and limited standardization continue to restrict routine adoption, and future AI-assisted analytical frameworks may be instrumental in addressing these issues and improving diagnostic reliability.

The rise of antifungal resistance—exemplified by multidrug-resistant Candida auris and azole-resistant Aspergillus fumigatus—has intensified the need to accelerate susceptibility testing. Conventional broth microdilution requires at least 24-48 hours following organism isolation to be completed, leaving a critical interval when empirical therapy may be inappropriate.

AI provides the possibility of inferring susceptibility directly from diagnostic data. At the single-cell level, ML that was applied to SCRS predicted fluconazole and amphotericin B susceptibility with 99% and 94% accuracy, respectively, in C. auris34. Delavy et al. (2020) used ML in predicting fluconazole resistance in Candida albicans with MALDI-TOF spectra, showing that spectral signatures carry information predictive of the resistance phenotype40. At a more system-level scale, Mutisya and Kanguha (2024) developed AntiMicro.ai, a web-based ML platform designed to predict antibacterial and antifungal susceptibility from clinical microbiological data, where an XGBoost model achieved AUCs of 0.990 and 0.978 for antibacterial and antifungal susceptibility prediction, respectively41.

The AI-based prediction of susceptibility is a promising route to faster therapeutic decisions, but recently developed models are trained on relatively limited datasets, so their results require extensive external validation before they can reliably guide treatment. Integration with routine laboratory workflows remains an important challenge.

Several multimodal AI systems have demonstrated improved performance where complementary clinical and imaging information is integrated. Drawing on complementary infor- mation, such models can capture complex disease patterns that are not apparent where datasets are analyzed in isolation, and they more closely mirror the integrative reasoning that characterizes real-world clinical decision-making. The com- bined CT models of Zhang et al. and Liu et al. and the multimodal keratitis model of Prajna et al. all illustrate this advantage16,17,24.

These observations demonstrate the development of in- tegrated diagnostic platforms that unify imaging, spectroscopic, genomic, and clinical data in a single AI framework. Such systems could improve accuracy, enable earlier intervention, and better reflect the multifactorial nature of fungal disease—although, as with individual modalities, prospective vali- dation and standardization must precede widespread clinical implementation.

12.1 Dataset scarcity and imbalance

The most pervasive constraint in this field is the shortage of large, well-annotated, and diverse fungal datasets. Most studies reviewed here were single-center and retrospective, with relatively small samples that limit statistical power and generalizability, and rarer but clinically important species are particularly underrepresented. The OpenFungi initiative is a welcome move toward standardized, publicly accessible benchmark datasets, but substantially larger multicenter collaborations will be required to meaningfully advance the field15.

12.2 External validation and prospective studies

External validation of data from other institutions and geographic settings remains the exception, not the rule. The pronounced geographic bias in the keratitis literature—where approximately 60% of studies focus on Chinese populations21—raises fundamental questions about whether reported per- formance generalizes across ethnicities, clinical practices, and healthcare systems. Prospective validation, as exemplified by Prajna et al. (2025)24, is uncommon yet indispensable for establishing genuine clinical utility.

12.3 Model explainability and clinical trust

Many deep-learning models are criticized for their "black box" character, a major barrier to clinical adoption: clinicians are understandably reluctant to base diagnostic and thera- peutic decisions on predictions that they cannot scrutinize. While some studies have used explainability techniques, such as Grad-CAM, SHapley Additive exPlanations (SHAP), and attention visualization23, the effective integration of inter- pretable AI into fungal diagnostic pipelines requires further work. Future systems should treat explainability as a first-class design requirement rather than part of a retrospective addition.

12.4 Regulatory and implementation barriers

The regulatory pathway for AI-based diagnostic devices in mycology is not yet well defined. As the frameworks of the US Food and Drug Administration, the European Union (CE-IVDR), and other bodies mature for the use of AI in radi- ology and ophthalmology, fungal-specific applications have attracted comparatively little regulatory attention. Integration with existing laboratory information systems and electronic health records will entail further technical and operational complexity.

12.5 Emerging opportunities

Several emerging trends nevertheless invite optimism. Foundation models that are pre-trained on large-scale bio- medical image corpora can mitigate data scarcity for task-specific fine-tuning, while federated learning provides a means of training shared models across multiple centers without requiring the exchange of sensitive patient data, thereby addressing both scarcity and privacy. Deploying AI on edge devices and mobile platforms is particularly promising for point-of-care diagnosis in resource-limited settings that tend to bear the greatest fungal disease burden. Additional avenues include large language models for use in clinical decision support, active-learning frameworks that prioritize the most informative samples for expert annotation, and continual-learning systems that are refined through ongoing clinical use. The eventual convergence of AI with rapid platforms such as SCRS and electronic noses could enable same-day, culture-free, species-level diagnosis with parallel susceptibility prediction—a prospect that would transform the manage- ment of conditions such as candidemia. A synthesis of the AI approaches that have been discussed in relation to various diagnostic modalities is presented in Table 1.

AI and ML have shown considerable promise across the full diagnostic spectrum of fungal infection. From microscopic image analysis and CT radiomics to ophthalmic and dermatological imaging, mass spectrometry, SCRS, biosensor-based VOC profiling, and mNGS bioinformatics, AI-driven methods frequently match or surpass the performance of human experts under controlled study conditions and the multi-modal integration of imaging, spectroscopic, genomic, and clinical data shows the most promising paradigm for the maximization of diagnostic performance. However, this field remains at an early translational stage. Large-scale, multicenter, prospective studies will be pivotal to overcoming the central obstacles of data scarcity, limited external validation, insufficient explainability, and an immature regulatory land- scape. While many AI-based tools continue to be in early development, their thoughtful integration into clinical work- flows could meaningfully improve the speed and accuracy of fungal diagnosis. Future progress will depend on multicenter collaboration, larger annotated datasets, and standardized evaluation frameworks to ensure reliable and equitable clinical implementation.