The architectures that we employed belong to the autoencoder (AE) family, a family of neural networks composed of an encoder and a decoder that learns to map inputs to outputs with minimal information loss through a low-dimensional latent space. For anomaly detection, this architecture works by reconstructing normal samples while failing to reconstruct outliers.

Many approaches in the literature focus on either fungi or bacteria classification, most commonly performing species-level classification within these two groups. Classification among different fungal genera, species, or higher-level classes that encapsulate multiple similar species (depending on the dataset) is most often addressed using CNN-based architectures, as demonstrated in works such as Bhimavarapu et al. (2024), Mohamed et al. (2024), Nawarathne and Kumari (2023), Rahman et al. (2023), Prommakhot and Srinonchat (2024). These studies show that CNN-based models can achieve strong performance on fungal image classification tasks. Transformer-based architectures have also been employed for this purpose and have achieved competitive results (Rahman et al., 2025; Gümüş, 2024; Ahmed and Haque, 2023). Similarly, CNN-based methods (Visitsattaponge et al., 2024; Poomrittigul et al., 2022), as well as transformer-based approaches (Prasad et al., 2025), have been proposed for bacterial classification tasks.

Research explicitly addressing bacterial versus fungal classification is less extensively explored in the literature; nevertheless, several relevant studies exist. One notable application is the differentiation between bacterial keratitis and fungal keratitis. In Redd et al. (2022), the authors evaluate five different CNN architectures, as well as a CNN ensemble, for classifying anterior segment eye photographs according to the etiological category of keratitis. Another approach is presented in Won et al. (2023), where a CNN with a ResNet (He et al., 2016) backbone is augmented with a Lesion Guiding Module and a Masked Adjusting Module, and qualitative analysis is performed using Grad-CAM visualizations. A multimodal approach is explored in Prajna et al. (2025), where both clinical data and images are used. The proposed model employs separate feature extractors for each modality, followed by a fusion mechanism prior to classification. The study reports that the vision-based model outperforms the clinical-data-only model and that the multimodal approach does not yield a significant improvement over the vision-only approach.

Studies focusing on bacteria and fungi classification from microscopic images can also be found in the literature. In Sroka-Oleksiak et al. (2025), the authors propose a method for identifying 14 bacterial species and 3 fungal species from Gram-stained microscopic images. Their pipeline includes cell segmentation using the Cellpose model, patch extraction, patch encoding, and aggregation prior to final classification. Separate classifiers are trained for bacterial and fungal samples. Other works, such as Treebupachatsakul and Poomrittigul (2020) and Zawadzki (2020), investigate multiclass classification of mixed bacterial and fungal microscopic images. In Treebupachatsakul and Poomrittigul (2020), a LeNet CNN is trained to distinguish between three bacterial species and one yeast species. In Zawadzki (2020), the authors evaluate multiple CNN architectures (Xception, ResNet, Inception-V3, MobileNetV2, and DenseNet201) on microscopic images stained using Congo Red and modified Gram stain across three datasets. They report accuracies of up to 100% on DS1 and 96% on DS2 using the Xception architecture.

The problem of classifying abscess images as bacterial or fungal is a binary classification problem. Denoting by C = { bacterial , fungal } the set of two targeted classes, the learning task is to estimate for an abscess image I the most probable class of C to which it belongs.

We formulate the problem as a one-class outlier detection problem in which the learning model is trained on only one class. The goal is to learn the main characteristics of the reference class and to detect as outliers those instances whose features differ significantly from the learned representation.

Let us consider a dataset of images of fungi and bacteria from abscesses. We consider the reference class (and denote it by c ref ) as the majority class (i.e. the one that has more available images), while the minority (less represented) class will be considered as the other class and will be denoted by c other . For instance, if more images of fungi from abscesses are available than those of bacteria from abscesses, then the reference class is the “fungal” one ( c ref = fungal) and the other class is the “bacterial” one ( c other = bacterial). Otherwise, if the fungal class is underrepresented in the dataset, then c ref = bacterial and c other = fungal.

OD-Abs model is trained only on the reference class c ref and learns the characteristics of this class by compressing and then recreating the images from the reference class. Theoretically, the model learns a decision boundary surrounding the images from the reference class. Thus, the model will be able to identify images that resemble the learned representation. During inference, new images that fall outside the decision boundary are classified as outliers (i.e. belonging to the c other class), as they exhibit significantly different characteristics than the representation of the reference class. Since OD-Abs is an encoder-decoder model, images from the c other class typically result in high reconstruction errors because the model has not learned to represent them.

More specifically, we would like to detect whether an abscess image I resembles images belonging to the reference class or is dissimilar to it and thus represents an anomaly. To determine whether the image I does not belong to the class c ref we compute the likelihood that the image is an anomaly of the reference class. This likelihood is expressed as the dissimilarity degree of the given image with respect to the structure of the class c ref encoded in a encoder-decoder architecture. Through their latent representations, the encoder-decoder architectures are used to uncover, through self-supervision, features and structural information from abscess images, which are relevant for distinguishing between fungi and bacteria.

Denoting by I m g the input space of abscess images, the classification of images of bacteria versus fungi from abscesses can be formalized as an approximation of a target function f : I m g → C that maps an abscess image to the “fungal” or “bacterial” class. Thus, the OD-Abs classification model will approximate an additional function p ref : I m g → [ 0 , 1 ] expressing the probability that an image I ∈ I m g belongs to the class c ref . Certainly, the probability p other that an image I ∈ I m g belongs to the class c other (i.e. it is an outlier) is obtained as p other ( I ) = 1 − p ref ( I ). Thus, after the approximation p ˜ ref ≈ p ref is learned from a training data consisting of images belonging to the majority (reference) class, an approximation of the target function f ˜ ≈ f will be simply computed as shown in equation (2). (2) f ˜ ( I ) = c ref . if p ref ( I ) ⩾ 0.5 , c other , otherwise , ∀ I ∈ I m g .

The overall OD-Abs methodology consists of three main stages. Figure 1 illustrates the two operational steps of the final classifier, namely training and thresholding, followed by classification of unseen images.

As illustrated in Fig. 1, our proposed OD-Abs approach uses an encoder-decoder model M that is trained through self-supervision on the images from the majority class (representing fungi or bacteria) and will learn to encode the structure of these images.

For the encoder-decoder model M, three variants are considered. The architectures of the three encoder-decoder models employed in our experiments, together with details regarding the training process, are described below.

As illustrated in Fig. 1, our proposed OD-Abs approach uses an encoder-decoder model M that is trained on a dataset of images belonging to the reference class C ref to encode information about the structure of the trained images. For training the model, 60 % of the images will be used, 20 % for the validation of the model, and the remaining 20 % will be used for testing.

The model M is trained through self-supervision to encode the structure of the majority class images, using an error function that penalizes the model; this function is expressed as the difference between the provided input image and the output image reconstructed through the decoder component.

Using the encoder-decoder model M selected following the evaluation introduced in Section 3.3 and trained as previously detailed, the OD-Abs classifier will produce an approximation p ref M of the probability function p ref : I m g → [ 0 , 1 ] expressing the probability that an abscess image I ∈ I m g belongs to the refernce class ( c ref ).

The learned function p ref M will be used by the OD-Abs classifier to predict if an abscess image belongs to the fungal or bacterial class; the probability that an image I belongs to the bacterial class is computed as p other M ( I ) = 1 − p ref M ( I ).

Let us denote by I ˜ M the output of the trained model M for an input image I (i.e. the reconstruction of I) and by RE ( I , I ˜ M ) the reconstruction error obtained for I ( L 1 distance between the input and the reconstruction). The probabilities p ref M ( I ) and p other M ( I ) estimated by the OD-Abs classifier using the encoder-decoder model M are computed as shown in equation (4). (4) p ref M ( I ) = τ M τ M + RE ( I , I ˜ M ) , p other M ( I ) = RE ( I , I ˜ M ) τ M + RE ( I , I ˜ M ) .

The idea behind computing the equation (4) is that the encoder-decoder model M can accurately reconstruct the images sampled from the same distribution as the training images (i.e. images of the reference class). Moreover, the model M should not be able to accurately reconstruct through its learned latent representation an input image that is structurally dissimilar to the training images, e.g. the images that belong to the minority class.

In equation (4), τ M is a hyperparameter of the OD-Abs classifier and represents the threshold applied on the reconstruction error of an input image I to decide if it belongs to the class c ref (i.e. if it is similar enough to the fungal class represented by the encoder-decoder model M). One observes from equation (4) that if RE ( I , I ˜ M ) < τ M , then p ref M ( I ) > 0.5; thus, I will be classified as belonging to the fungal class c ref . Otherwise, if RE ( I , I ˜ M ) ⩾ τ M , then p ref M ( I ) ⩽ 0.5, then I will be considered dissimilar to the class represented by the model M and classified as belonging to the bacterial class c other . It should also be noted that if RE ( I , I ˜ M ) = 0, then p ref M ( I ) = 1, while if RE ( I , I ˜ M ) = τ M , then p ref M ( I ) = 0.5. Moreover, the value of p ref M ( I ) decreases as RE ( I , I ˜ M ) increases.

Two variants are considered for computing the hyperparameter τ after completing the training of OD-Abs. V1

For a specific encoder-decoder model M we are computing the mean ( μ ( M )) and the standard deviation ( σ ( M )) of the reconstruction errors RE ( I , I ˜ M ) computed for all the images from the training dataset. If D is the training dataset and n represents the number of images from the training set ( n = | D |), then μ ( M ) = ∑ I ∈ D RE ( I , I ˜ M ) n and σ ( M ) = ∑ I ∈ D ( RE ( I , I ˜ M ) − μ ( M ) ) 2 n .

The first method for estimating the threshold τ is to compute it as the mean of the reconstruction errors of the training images plus twice the standard deviation of the errors, i.e. τ M = μ ( M ) + 2 σ ( M ). The intuition behind this computation is that the rule “mean plus twice the standard deviation of the values from a population” is important in statistical distributions, particularly in the normal distribution. Even though the maximum of a population is not necessarily located at the μ ± 2 σ, this formula gives an estimate of a good enough “upper bound” of the population values, as in many normal distributions, most values lie within μ ± 2 σ.

After training, the OD-Abs classifier is tested on both images of fungi and bacteria from abscess to assess its predictive performance.

As previously described in Section 3.1, based on the number of abscess images available for the “fungal” and “bacterial” classes, the majority class is considered the reference class and used for training the OCC model. Besides the dataset of abscess images of the reference class used for training, validation, and testing the OD-Abs model, a dataset of abscess images of the minority class will be used for testing.

As shown in Section 3.4, the training dataset is randomly split into training-validation-testing. Due to the randomness of the splitting process, a cross-validation testing methodology is used. More specifically, the data set of images of the majority class is randomly split into training-validation-testing. For a specific split, the OD-Abs classifier will be tested on a dataset obtained by combining 20 % of the training images that were not used for training and validation, and the dataset of images of the minority class.

For measuring the performance of OD-Abs on a testing set, the following evaluation measures are computed (Gu et al., 2009), considering the “fungal” class as the negative one and the “bacterial” class as the positive one: precision ( Prec = T P T P + F P ), sensitivity (recall, or true positive rate, Sens = T P T P + F N ), specificity (true negative rate, Spec = T N T N + F P ), F1 score ( F1 = 2 · P r e c · S e n s P r e c + S e n s ) and Area Under the ROC Curve ( AUC = S p e c + S e n s 2 ).

We note that for computing the AUC measure, we consider the case of classifiers that provide the output class directly (Fawcett, 2006). In this case, the ROC (Receiver Operating Characteristic) curve is a quadrilateral obtained by linking the points ( 0 , 0 ) and ( 1 , 1 ) to the ( 1 − Spec, Sens) point. According to the literature, AUC is one of the most effective evaluation metrics for assessing the performance of imbalanced classifiers (Fawcett, 2006). For each of the previously described performance metrics, the performance of OD-Abs averaged over the five runs is reported together with the 95% confidence interval (CI) (Brown et al., 2001) of the mean.

For classifying an input abscess image as having a fungal or bacterial origin, the trained OD-Abs model performs the steps described in Section 3.4 and summarized in Algorithm 1.

This section describes the datasets used in the experiments: the labelled dataset D f containing images of fungi from abscesses and the labelled dataset D b containing images of bacteria from abscesses.

For D f we used the DeFungi dataset (Sopo et al., 2021), a publicly available repository of 9114 manually labelled microscopic images. These images were obtained from direct mycological examinations of fungal infections and processed into 500 × 500 pixel patches. The dataset is organized into five classes corresponding to morphological patterns observed under microscopy: Tortuous septate hyaline hyphae (TSH), Beaded arthroconidial septate hyaline hyphae (BASH), Groups or mosaics of arthroconidia (GMA), Septate hyaline hyphae with chlamydioconidia (SHC), and Broad brown hyphae (BBH).

For D b we employed both the DIBaS (Digital Images of Bacterial Species) (Zieliński et al., 2017) dataset and a subset of the DeepBacs (Spahn et al., 2022) image collection, more specifically the raw images from the mixed segmentation dataset. The DIBaS dataset contains approximately 640 bacteria Gram-stained microscopic images covering 32 bacterial species. The DeepBacs dataset that we used contains 190 brightfield and fluorescence images covering three species: S. aureus, E. coli and B. subtilis.

Figure 2 comparatively illustrates a fungi image (the left side of the image) and a bacteria image (the right side of the image). One observes that fungi appear as filamentous hyphae (marked with h in the left-side image), comprised of long, tube-like strands with rounded segments identified as possible spores or conidia (marked with s in left-side image), in contrast to the bacterial cells that look like clustered spherical cells typical of Gram-positive cocci (marked with arrows in the right-side image).

We note that in our experiment the bacterial class is underrepresented, and thus the reference class is the majority “fungal” class ( c ref = D f), while the minority class is the “bacterial” class ( c other = D b).

According to the methodology introduced in Section 3, the OD-Abs classifier is trained on D f and learns a decision boundary by representing the characteristics of the fungal class. At the query time, the images that resemble the learned representation are predicted as belonging to the “fungal” class; the images that fall outside the decision boundary (i.e. deviate significantly from the learned representation) are considered outliers and predicted as “bacterial”.

The performance of the three encoder-decoder models introduced in Section 3.3 is comparatively analysed to decide the best-performing encoder-decoder model to be integrated in the OD-Abs one-class classifier.

The dataset containing images of fungi from abscesses ( D f) and the dataset of images of bacteria from abscesses ( D b) will be used for evaluating the performance of the three encoder-decoder models proposed. From a one-class outlier detection perspective, the best model is considered the one that provides an accurate reconstruction of images of fungi but fails to reconstruct images of bacteria.

Table 2 presents the values of the metrics computed on D f and D b. The best values obtained for the metrics are highlighted. We note that we search for higher similarities between the fungal images and their reconstruction (lower MSE_Df, higher SSIM_Df and SNR_Df), while lower similarities between the bacterial images and their reconstruction (higher MSE_Db, lower SSIM_Db and SNR_Db). In order to evaluate the model’s ability to distinguish between the two classes, we utilize the Sensitivity Index (SI), defined as d ′ = | μ D f − μ D b | 1 2 ( σ D f 2 + σ D b 2 ) . A higher SI value indicates superior class separability. The best values for the evaluation metrics are highlighted.

Table 2 reveals a trade-off between reconstruction fidelity and separation capability across the three models. While CAE achieves the best reconstruction metrics on the target fungal data (lowest MSE_Df of 0.002 and highest SNR_Df), it fails to sufficiently differentiate the bacterial samples, shown by the low MSE_Db and the lowest Sensitivity Index ( d ′ = 2.156). This suggests that while CAE is a good compressor, it lacks the necessary regularizing constraints to distinguish between the two biological classes effectively.

In contrast, the variational approaches provide better separability. Although the standard VAE offers a balanced performance, the β-VAE model proves to be the superior version. It yields the highest error and lowest SNR when reconstructing bacterial images (MSE D b = 0.078) while maintaining comparable performance on the fungal class. Most importantly, β-VAE achieves the highest Sensitivity Index ( d ′ = 3.205), indicating the largest statistical distance between the normal and anomalous distributions. Therefore, based on the maximization of class separability, β-VAE is selected as the most suitable model for the proposed anomaly detection framework (see Fig. 3).

As shown in Section 3.3, the β-VAE model was empirically found to be the best encoder-decoder model in preserving the structural characteristics of images of fungal abscesses, while failing to reconstruct images of bacterial abscesses. Thus, we hypothesize that the OD-Abs classifier with the β-VAE model integrated can accurately encode the fungal class and uncover hidden patterns and structural information useful to effectively determine whether images of abscesses have a fungal or bacterial origin.

To complete the answer to RQ1, the performance of the OD-Abs classifier incorporating the β-VAE encoder-decoder is evaluated on the datasets of images of fungi and bacteria from abscesses described in Section 4.1. For the hyperparameter τ M (the decision threshold) both variants (V1, V2) described in Section 3.4 are used. Table 3 presents the experimental results obtained following the testing methodology introduced in Section 3.5. The values for the performance metrics are averaged over five runs, and the 95% CI of the mean values is reported. The best values for the performance metrics are highlighted.

Analysing the performance metric values from Table 3, one observes that, in terms of AUC, both variants for selecting the threshold τ M yield high performances with a clear advantage for V2. We consider that variant V2 provides a better balance between the true positive rate (TPRate, Sens) and true negative rate (TNRate, Spec). We note that while V1 provides a near-perfect TPRate, it exhibits a lower capability to recognize fungal images from abscesses (lower Spec) compared to V2, meaning more fungal images are incorrectly flagged as bacterial/anomalous. Moreover, compared to V1, V2 provides a much lower false alarm ratio (FAR, the proportion of images incorrectly predicted as “bacterial” from the total number of predicted bacterial images), indicated by its superior Prec value (0.967 vs 0.764). A low FAR is desirable, as it indicates that fewer fungal images are incorrectly flagged as bacterial. This is relevant in practice, since fungal infections are generally less frequent than bacterial infections and should therefore not be overlooked.

To better highlight the good performance of our proposed OD-Abs one-class classification model, we performed an additional comparison to binary classification models. However, this comparison requires important contextualization regarding the learning paradigms employed. Obviously, it is expected that the binary classifiers achieve better predictive performance than the OCC ones, since the former are trained on images from both classes to learn the decision boundary between them, while the latter receive samples from only one class and learn a boundary surrounding the images from the reference class. However, the aim is to see how close or far the performance of OD-Abs is from the baseline binary classifiers.

For comparison, we selected the architectures most commonly applied to the related tasks discussed in Section 2.2, such as species-level classification, macroscopic binary classification, and multiclass microorganism classification. Consequently, we compare our approach with baseline CNN and Transformer-based architectures, specifically ResNet50 (He et al., 2016), VGG16 (Simonyan and Zisserman, 2015), and a base Vision Transformer (ViT) (Dosovitskiy et al., 2021). These models were initialized with pre-trained weights and fine-tuned for five epochs on our fungal and bacterial datasets.

The supervised architectures achieved near-perfect separation, with ResNet50 reaching an AUC of 1.0000, followed closely by ViT (0.9993) and VGG16 (0.9976), surpassing the OD-Abs model. Our OD-Abs one-class classifier is outperformed by the baseline binary classifiers with at most 3.3 % in terms of Prec and F1, 3.2 % in terms of Sens, and 2 % in terms of AUC. We have to remark that in terms of Spec, OD-Abs outperformed VGG16 by 0.4 %, and ViT by 3.8 %, having nearly the same performance as ResNet50 (with only 0.8 % lower). Thus, our model provides lower misclassification rates for the images of fungi from abscesses (i.e. false positive rates) than VGG16 and ViT, being able to better distinguish the “fungal” class from the “bacterial” class. In this context, we consider the performance achieved by our OD-Abs model notable.

It has to be noted that the baseline supervised classification models benefit from a significant advantage: supervised training, allowing them to explicitly learn the decision boundary between fungal and bacterial features. In contrast, OD-Abs operates in a one-class setting, training exclusively on fungal images without exposure to bacterial samples during training. The fact that OD-Abs achieves a high AUC and F1 scores and a performance very close to the performance of binary classifiers, even outperforming them in terms of false positive rate, demonstrates the robustness of the one-class hypothesis. It suggests that the structural pattern of fungi is distinct enough that a model can learn to reject bacterial anomalies effectively without ever seeing them.

To conclude, the proposed OD-Abs model succeeds in ensuring a very good balance between preserving the structure of the images from the reference “fungal” class and uncovering the underlying structural relationships within them, and successfully recognizing the “bacterial” class as well. However, the gap in AUC between our custom model and the pre-trained supervised baselines implies that our current architecture may miss fine-grained texture representations.