Automated Detection of Waterborne Pathogens in Aquaculture Using an Enhanced Swin-Transformer Model

Waterborne pathogens are microorganisms in aquatic environments and can cause diseases in fish and other aquatic organisms. Aquaculture systems, which involve farming aquatic organisms for food or other purposes, can provide an ideal environment for the growth and spread of these pathogens [1]. It may put customers' health in danger as well as cause aquaculture producers to suffer large financial losses. Aquaculture systems can quickly and effectively detect the presence of these bacteria by using automated waterborne pathogen detection [2]. This process involves the use of advanced technology and equipment to analyze water samples and detect the presence of pathogens automatically. The first step in automated detection is the sampling process, where water samples are collected from different areas of the aquaculture system [3]. These samples are then filtered to isolate the microorganisms present in the water. The filtration can be done manually or through automated equipment. The samples that have been filtered are subsequently exposed to different techniques for detection, including enzyme-linked immunosorbent assay (ELISA), quantitative PCR (qPCR), and polymerase chain reaction (PCR) [4]. To locate and measure the pathogen in the sample, these techniques employ certain markers or antibodies. Automated detection offers several benefits, chief among them being its speedy sample processing, which allows for an earlier identification of possible pathogen contamination. It is essential in aquaculture systems, where a delay in detection could lead to significant economic losses. Automated detection reduces the risk of human error and contamination, as machines carry out the entire process under controlled conditions. It minimizes the chances of false results and ensures greater accuracy than traditional detection methods. Automated detection is also beneficial in terms of cost and efficiency. It requires less manual labor and consumables, making it a cost-effective method for regularly monitoring water quality in aquaculture systems. It also provides real-time results, which can be easily interpreted and analyzed, allowing for timely and appropriate actions to be taken in case of a pathogen outbreak [5].

The automated detection of waterborne pathogens in aquaculture systems is a crucial process for ensuring the health and safety of aquatic organisms and maintaining the productivity of the aquaculture industry. It allows for the efficient and accurate identification of pathogens, essential for disease control and prevention. As technology advances, automated detection methods are expected to become more widespread, leading to improved management and sustainability of aquaculture systems. Aquaculture, known as fish farming, cultivates and harvests fish, shellfish, and seaweed in controlled marine or freshwater environments. With the increasing demand for seafood, aquaculture has become an important industry worldwide. However, the rapid expansion of aquaculture has also led to several challenges, including the risks associated with waterborne pathogens. These microscopic organisms can cause diseases in aquatic animals, leading to significant economic losses for aquaculture producers. One way to mitigate this risk is through automated detection systems. These systems use advanced technologies such as DNA-based sensors, spectroscopy, and biosensors to detect the presence of pathogens in water samples. While this technology holds great potential, practical issues must be addressed for its effective implementation in aquaculture systems [6]. The main technical challenge with automated detection systems is ensuring their accuracy and reliability. These systems rely on complex algorithms and data analysis, making it essential to validate and calibrate them for accurate results continually. As water samples may contain a mix of pathogens, these systems must accurately identify and differentiate between different species. Any detection error could lead to false positives or negatives, which can seriously affect the aquaculture industry [7]. Another issue is the cost of these automated systems, which can be a significant barrier to their widespread adoption. Many technologies require expensive equipment and skilled personnel for operation and maintenance. It makes it difficult for small-scale aquaculture producers to invest in these systems, limiting their access to this technology. Additionally, the analysis of water samples can be time-consuming, and there is a need for rapid detection to prevent the spread of diseases and minimize economic losses effectively [8]. Further, developing and standardizing protocols for automated detection systems are essential to ensure consistent and reliable results across different laboratories and systems. Another major challenge is the diversity of waterborne pathogens in aquaculture systems. These systems can house a variety of species, each with its own unique set of pathogens, making it challenging to design universal detection systems [9]. Additionally, environmental factors, such as water temperature, pH levels, and nutrient levels, can influence the growth and presence of these pathogens, further complicating the detection process. In conclusion, while automated detection systems hold immense potential in addressing the risks of waterborne pathogens in aquaculture systems, several technical challenges must be addressed [10]. These include ensuring the accuracy and reliability of the system, managing the high costs involved, developing standardized protocols, and accounting for the diversity of pathogens in aquaculture systems [11]. Addressing these issues will be crucial in successfully adopting and implementing automated detection systems for the sustainable growth of the aquaculture industry. The main contribution of the research has the following,

• The Swin-Transformer uses a deep learning algorithm to detect and classify waterborne pathogens with high accuracy. This allows for automated and reliable detection of even low levels of pathogens in aquaculture systems.

• The Swin-Transformer is designed to efficiently process large datasets, making it an ideal tool for monitoring water quality in aquaculture systems. It can quickly analyze complex data and provide real-time results, allowing prompt action in case of pathogen detection.

• The proposed Swin Transformer can detect multiple types of pathogens simultaneously. This is essential for the aquaculture industry, which often faces the challenge of multiple pathogens in a single system.

• The Swin-Transformer can accurately identify and differentiate between different pathogen types, providing a comprehensive understanding of the water quality in aquaculture systems.

This study addresses these limitations by proposing an enhanced Swin-Transformer model that directly tackles three specific gaps in the current research landscape:

(1) Limited capacity of CNN-based models to capture long-range dependencies and contextual information—essential for distinguishing between visually similar pathogen features in complex scenes.

(2) Inefficiency and scalability issues in existing two-stage object detectors when applied to high-throughput image analysis in aquaculture systems.

(3) Lack of unified models capable of simultaneously detecting and classifying multiple pathogens with high accuracy and speed under variable environmental and image quality conditions.

To bridge these gaps, we propose a novel architecture that combines the local feature extraction power of CNNs with the global reasoning ability of the Swin-Transformer’s hierarchical attention mechanism. The enhanced model is designed to support real-time, multi-pathogen detection with high accuracy, leveraging self-attention, shifted windows, and pre-training to improve generalization and robustness.

1.jpg

By calculating similarity scores across all positions, the self-attention mechanism captures global spatial relationships, in contrast to convolution operators that extract local correlations.

$b_v=p_v=o_t=\sigma\left(H_v+X_{t-1}\right)$                (1)

$X_v=o_v \circ \tanh \left(C_t\right)$                 (2)

For model training, Swingier introduces some improvements, such as data augmentation, adversarial training, and progressive learning.

$R(\theta)=\frac{1}{M} \sum_{b=1}^M \sqrt{\left({Modl}\left(I_R^b, \theta\right)=I_X^b\right)^2}$              (3)

Eq. (2) can be resolved by applying the optimization transfer method and forward-backward splitting. The objection function is divided into two terms using the FBS algorithm.

$l^{k y}=h^{y-1}-\alpha \beta \nabla L\left(h^{y-1}\right)$                (4)

$h^y=\underset{x}{\arg \max } \,L(h \mid k)-\frac{1}{2 \alpha}\left\|h-l^y\right\|^2$                 (5)

Our suggestion was to substitute the gradient descent update in the problem with a residual Swin Transformer based regularize (RSTR).

$l^y=R S T R\left(h^{k-1}\right)$                (6)

Quantitative comparisons against images with high dosage labels were carried out to evaluate the quality of the reconstruction. The normalisation of references and reconstructed pictures was set to a maximum value of 1. The object Detection based on Swin-Transformer utilizes a single-stage architecture with a hierarchical Transformer structure, efficient attention mechanisms, and a pre-trained backbone model to detect and classify objects in an image or a video. Its ability to handle typical and abnormal data and use an embedding function makes it a powerful and versatile object detection framework.

3.1 Feature extraction

An artificial neural network known as a Multilayer Perceptron (MLP) is a widely used model that consists of several layers of connected nodes or units, each with a specific number of neurons. An MLP's W-MSA (Weighted-Majority Summing Activation) activation function calculates a neuron's output by adding up all of the weighted inputs from the layer before it. The four primary phases of an MLP's operation with W-MSA are the input layer, hidden layers, output layer, and backpropagation. The feature extraction is displayed in the subsequent Figure 2.

(1) Input layer: An MLP's input layer is responsible for receiving its input data, which is often expressed as a vector. A corresponding input value for every neuron in the input layer is transmitted to the subsequent layer.

(2) Hidden layers: Most of the processing in an MLP is done by the hidden layers. There are numerous neurons in each buried layer, and each neuron in the layer above receives input from all of the neurons in the layer below it. Next, the W-MSA activation function is applied to every neuron, which produces its output by applying a non-linear transformation and computing the weighted sum of its inputs.

(3) Output layer: The output layer is the last in the network, and it uses the inputs from the hidden layers to get the final classification or prediction. Each neuron's output is likewise calculated using the W-MSA activation function, and the error is determined by comparing the actual output to the expected output.

(4) Backpropagation: Using the computed error as a guide, this method updates the MLP's weights. Using an optimisation approach like gradient descent, the weights are modified by the mistake that is propagated from the output layer back to the hidden layers. With the ultimate objective of reducing error and raising network accuracy, this procedure is carried out for every input in the training set.

2.jpg

Figure 2. Feature extraction process

The Transformer encoder is divided into four primary phases and consists of L layers. The image is first divided into a series of non-overlapping patches.

$x_i=w_{ {patch }} \times patch _i$              (7)

Second, to encapsulate the spatial information of the image, learnable position embeddings are added to each patch.

$x_i=x_i \times p o s_i$               (8)

A user with more processing power might try to recreate an image of excellent quality, while a user with less processing power would try to produce an image that is rough and lacks fine details.

$s_{I, k}=G_\alpha\left(B / d_y\right)$                 (9)

The encoded symbols are displayed as, and the channel encoder's parameter set is designated as β.

$H_I=D_\beta\left(s_I\right) \in \mathbb{R}^k$                 (10)

MSE loss can be used to calculate the difference between the original image and the one that user k rebuilt, as demonstrated below:

$c\left(B, \widehat{B}_y\right)=\frac{1}{n} \sum_{b=1}^m\left(B_i-\widehat{B}_i\right)^2$               (11)

where, the product of the image's height, width, and number of channels equals its size, denoted by n. Because it adds non-linearity to the network and enables it to recognise intricate patterns and relationships in the data, the W-MSA activation function is essential to the functioning of an MLP. The network can also learn hierarchical representations of the data by utilising many hidden layers with W-MSA, which enhances network performance even more. In general, the amalgamation of MLP and W-MSA enables the pragmatic training of profound neural networks, rendering it an efficacious instrument across diverse domains, including but not limited to image and audio identification, natural language processing, and predictive modelling.

3.2 Classification

Input is the initial stage of the operation, during which the input image is fed into the system. Depending on the application, the input can either be a single image or a batch of images. Once the input is received, it is passed through the backbone network. The backbone network is responsible for extracting meaningful features from the input image. It is done through convolutional and pooling operations, which learn to detect low-level features like edges, textures, and shapes. After the backbone network, the image is passed through the C3SwinTR*6 module. The Swin Transformer classification module has shown in the following Figure 3.

It is a modification of the famous Swin Transformer architecture and stands for Cross-Crop Cross-Stage Swin Transformer with a depth of 6. The C3SwinTR*6 module takes the features extracted by the backbone network and applies the Transformer network. This network uses attention mechanisms to capture long-range dependencies and allows feature fusion across multiple scales. It enables the network to handle large images and effectively capture global and local features. The Swin Transformer block is the fundamental building block of the C3SwinTR*6 module. It is composed of several layers, with a feed-forward network, a layer normalisation layer, and a self-attention layer in each layer. The network's comprehension of the connections between various visual components is aided by the self-attention layer. The non-linear changes made to the features are picked up by the feed-forward network. By standardising each layer's output, the layer normalisation layer aids in maintaining stability during the learning process. The features from the backbone network are then fused with the C3SwinTR*6 module manufacturing through the use of the Fusion Concat operation. This operation concatenates the two sets of features, preserving the spatial information from both sources. It allows the network to combine local and global features, thus improving performance effectively. After the fusion operation, the features are passed through the upsampling layer. This layer increases the resolution of the features, allowing the network to learn more detailed and fine-grained features. It is essential for tasks that require precise localization, such as object detection. In this work, we created two three-branch feature fusion blocks, utilising the Basic-Block block that is suggested as a more flexible way to form networks.

3.jpg

$f_{ {Msmt }}=P_4\left(P_1(h)+P_2(h)+P_3(h)\right)$             (12)

where, the function of the Basic-block module is indicated by F (•). x indicates the module Most's input, and x is its output. The multi-head attention mechanism in self-attention serves as the inspiration for this design. A dynamic weight technique like this directs the model to concentrate more on enhancing performance during the plateau period, which then increases the model's effectiveness in the other phases.

$R_{N V R}=\gamma . L_{N V R}(h)$                (13)

Conversely, the combined loss function can be expressed as follows by combining the MTL loss and knowledge distillation loss.

$R_{ {comp }}=(1-\alpha) L_{N V R}+\alpha R_{{dis }}$,                (14)

where, the hyperparameter for the balance loss is α. For our studies, we set α = 0.1 after the original loss and feature loss magnitudes are harmonised. We attach more weight to it. A dynamic weight technique like this directs the model to concentrate more on enhancing performance during the plateau period, which then increases the model's effectiveness in the other phases.

$R_{N V R}=\gamma . R_{N V R}(h)$                  (15)

With this action, the location information in one direction is preserved while the channel dependencies in the other direction are captured by the attention map.

$e=\delta\left({Conv}_1\left(\left[p^x, p^w\right]\right)\right)$                  (16)

Indicates the spatial dimension concatenation operation, the 1×1 convolution operation is represented by Conv1(•), and the nonlinear activation function is represented by δ (•). By optimising channel output to 256 and fine-tuning the weighted features, this procedure essentially reduces parameters. The CAM module's output can be found as

$k_d(b, a)={Conv}_1\left(h_d(i, a) \times o_d^x(b) \times o_d^w(a)\right)$                 (17)

Figure 2's "Attention weighted fusion" section shows the layout of our suggested AWF module. Our module efficiently captures inter-channel correlations at various sizes by implementing the suggested CAM at every level of the feature pyramid. Next, an approximate binary map is computed using the following formula, which takes into account the pixel relationship between the text regions and the text boundaries.

$\tilde{I}_{b, a},=\frac{1}{1+g^{-y}\left(P_{b, a}-V_{b, a}\right)}$                  (18)

where, $P_{b, a}$ and $V_{b, a}$ denote the predicted pixel values from the segmentation network at coordinate (b, a) of the probability map and threshold map, respectively. g is a scaling factor. y controls the sensitivity of the function. During training, the total loss function L is decomposed into three components as follows:

$R=R_{i b}+\lambda_1 R_{f l}+\lambda_2 R_{t h}$                (19)

Because it is less dependent on the class probability prediction of pixels and more focused on the degree of overlap between prediction results and the genuine labels, it has an advantage when handling issues like sample imbalance and boundary-blurring.

$R_{i b}=1-\frac{2 \sum_{i=1}^M\left(h_i \times k_b\right)}{\sum_{b=1}^M h_b+\sum_{b=1}^M k_b+\varepsilon}$                   (20)

where, h_i and k_b represent the pixel's genuine label value and prediction in the approximate binary map. To prevent division by zero and lessen overfitting, the smoothing item denoted by ε is utilised. The features are passed through the Neck module, the Neck block. This module takes the up-sampled features and further refines them using a combination of convolution and up-sampling operations. It helps to remove any noise or artifacts in the features and prepares them for the final output. Overall, the combination of these operations in the network enables it to effectively extract features at different scales, fuse them, and then refine them for the final output. It allows the network to perform state-of-the-art object detection and image classification tasks. From image input to final output, each operation plays a crucial role in the overall functioning of the network. The proposed algorithm has been shown in the following,

The proposed algorithm for pathogen detection using the Swin Transformer is designed to identify harmful pathogens in a collection of images by employing deep learning and mathematical preprocessing. The input to the algorithm is a set of images $I=\left\{I_1, I_2, \ldots \ldots \ldots I_n\right\}$, where each $I_i$ represents an individual image sample. Additionally, M denotes the pre-trained Swin Transformer model, a hierarchical vision transformer that processes images in smaller patches using shifted windows, enabling efficient feature extraction and robust predictions. The output is a set $D R=\left\{d r_1, d r_{2, \ldots \ldots \ldots} d r_{\mathrm{n}}\right\}$ where each dr_i corresponds to the detection result for $I_i$ indicating the presence or absence of pathogens. The algorithm begins by importing the required libraries (T) and loading the Swin Transformer model (M) in evaluation mode. This ensures the model operates strictly for inference, disabling any gradient updates that occur during training. Each input image undergoes preprocessing via a function PPI(I), which involves several mathematical steps. First, the image is resized to a fixed dimension (W, H) where W is the width, and H is the height, ensuring compatibility with the model's input requirements. Then, pixel values of the resized image are normalized using the formula I_normalized=$\frac{I_{ {resized }}-\mu}{\sigma}$ where μ is the mean, and σ is the standard deviation of pixel intensities. This normalization standardizes the image, aligning it with the distribution used during model training. The normalized image is then converted into a tensor, a multi-dimensional array format that deep learning models like Swin Transformer process efficiently.

The preprocessed image tensor is passed through the Swin Transformer model (M), which outputs predictions patho, identifying the pathogens present. This process is repeated for each image path p_i in the list of image paths $\left\{P_1, P_2, \ldots \ldots \ldots P_n\right\}$. For each image, the detection result dr_i= DDetectPathogens(PP_I(_Pi)) is computed and added to the set DR. If any dr_icorresponds to a harmful pathogen, an alert is generated using $Alert\leftarrow GenerateAlert$ (dr_i), signaling potential risks. Finally, the algorithm returns the complete set of detection results DR. The Swin Transformer’s hierarchical feature extraction and shifted window attention mechanism make it particularly effective for analyzing complex visual patterns in images. The mathematical steps of resizing, normalization, and tensor conversion ensure the input is optimally prepared for the model, while evaluation mode ensures efficient and accurate inference. This comprehensive process makes the algorithm robust and scalable for real-world pathogen detection scenarios.

The proposed Swin Transformer model is a robust and efficient tool for the automated detection of waterborne pathogens in aquaculture systems. By utilizing transformer-based self-attention mechanisms and the Swin Transformer architecture, the model effectively identifies harmful pathogens, even at low contamination levels. The self-attention mechanism captures long-range dependencies and intricate patterns, enhancing both accuracy and reliability. The Swin Transformer’s architecture processes smaller data windows independently before aggregation, improving computational efficiency and scalability, which makes it suitable for real-time pathogen detection in dynamic aquaculture environments. The model achieves impressive performance metrics, including 94.44% accuracy, 94.75% recall, and a 0.87 Matthew's correlation coefficient, showcasing its effectiveness in pathogen detection. The incorporation of innovative positional embedding techniques enables the model to capture critical spatial and temporal relationships, essential for monitoring pathogen concentration fluctuations over time. While the model shows significant promise, future improvements could consider the practical constraints of real-world deployment, such as computational cost and energy efficiency. Optimizing the model for edge devices while addressing these constraints would enhance its feasibility in remote aquaculture settings, enabling faster, more efficient detection with minimal resource consumption. Additionally, integrating multimodal data sources, such as environmental parameters and sensor-based readings, could further improve the model’s predictive capabilities in diverse aquaculture environments.