BioSwarmNet: A Revolutionary Approach to Brain Tumour Detection Using Fractional Order Differential Particle Swarm Optimisation and Recurrent Neural Networks

As of 2023, cancer statistics reveal a significant toll on public health, with an estimated 11,020 deaths attributed to brain and other nervous system cancers in males and 7,970 in females. Among children, brain tumors, including benign and borderline malignant types, account for 26% and 21% of diagnoses, followed by lymphoma at 12% and 19%. Excluding benign and borderline malignant brain tumors, which boast a 5-year relative survival rate of over 97% in children and adolescents, mortality rates remain concerning. Brain tumours are a significant and urgent healthcare concern that necessitates an accurate and timely diagnosis [1-3]. These abnormal cell growths within the brain can manifest in a variety of ways, posing a serious risk to an individual's health [4]. The importance of detecting and treating brain tumours stems from their proclivity to evolve rapidly, frequently resulting in severe neurological complications and even life-threatening conditions [5-7]. The prevalence of brain tumours is not only a medical issue, but also a societal one, as the consequences can be devastating for patients and their families [8].

Brain tumours can develop as a result of a variety of factors, including genetic predisposition and environmental influences [9]. While the exact causes are unknown, certain risk factors have been identified, including radiation exposure, a family history of brain tumours, and certain genetic syndromes [10]. Despite advances in medical science, the exact aetiology of brain tumours is frequently unknown, making early detection and diagnosis critical [11].

The ability of brain tumours to grow and infiltrate brain tissues, resulting in a variety of neurological symptoms, characterizes their evolution [12]. Headaches, seizures, changes in cognitive function, and motor deficits are all possible symptoms. Brain tumours can progress quickly, exacerbating these symptoms and necessitating immediate medical attention. In some cases, a tumour can grow to such a size or location that it causes a sudden neurological crisis, necessitating emergency treatment [13, 14].

The analysis identifies several shortcomings in current brain tumor detection methods, including low sensitivity, interpretability issues, and a reliance on meticulous parameter tuning. Existing approaches, such as optimization techniques and deep learning models, frequently struggle to capture intricate patterns and textures in brain tumor images, resulting in suboptimal performance and limiting high-level abstraction. Furthermore, while some models achieve commendable accuracy, they may be insufficiently robust and fail to meet current standards. BioSwarmNet addresses these challenges by integrating bio-inspired optimization techniques with Recurrent Neural Networks (RNNs), leveraging their collective capabilities to enhance accuracy, sensitivity, and interpretability in brain tumor detection.

Our research addresses the critical need for accurate and timely detection and diagnosis of brain tumours. Given the urgency of brain tumour cases, our contribution is focused on the development of an advanced medical image analysis system based on the novel "BioSwarmNet" model. This system is intended to improve the accuracy and efficiency of brain tumour identification significantly. Our research aims to provide a reliable tool for healthcare professionals by leveraging the renowned BRATS dataset and harnessing swarm intelligence and deep learning. This tool can help with the early detection and diagnosis of brain tumours, which can improve patient outcomes and reduce the burden on both patients and the healthcare system. In the following sections, we will go over our methodology in detail and present compelling evidence of our approach's efficacy in addressing this pressing healthcare challenge.

The paper is divided into five major sections: The introduction establishes the context by emphasizing the critical need for early and precise brain tumour detection, introducing our advanced medical image analysis system based on the "BioSwarmNet" model, and outlining the paper's structure. We provide context in the second section with a comprehensive literature review that summarizes recent research in brain tumour detection and its limitations. The third section delves into the architecture and operation of our proposed system, emphasizing the importance of each step and introducing the unique "BioSwarmNet" model, which combines Fractional Order Differential Particle Swarm Optimisation (FODPSO) and Recurrent Neural Networks (RNNs). The fourth section presents our research's findings and analysis, highlighting the impressive results obtained when applying our approach to the BRATS dataset, including superior accuracy, sensitivity, and specificity when compared to previous works. Finally, the paper summarizes our contributions, highlighting the transformative potential of "BioSwarmNet" in brain tumour detection, and providing comments on broader implications in medical image analysis and healthcare. The paper is supplemented by a references section that includes citations for further research on the topic.

The proposed system diagram is depicted in Figure 1 based on the methodology described in the text. The workflow starts with user interaction, which allows the user to select an image from the BRATS dataset [24]. A series of pre-processing steps are then applied to the selected image. To begin, it is resized to a standard 256x256 pixel dimension to ensure consistency. The luminance method is used to convert color images to grayscale [25]. A Wiener filter in the frequency domain is used to reduce noise. The power spectral density (PSD) and noise PSD of the image are used to calculate this filter, which results in a filtered image [26]. Following that, an Adaptive Histogram Equalization (AHE) step improves image quality, which is followed by the use of the Discrete Wavelet Transform (DWT) with the Haar wavelet [27]. To reconstruct the image, the DWT coefficients are combined. Finally, features from the Grey Level Co-occurrence Matrix (GLCM) are extracted, and the image is classified using a pretrained neural network, leading to tailored actions based on the classification outcome, such as tumour localization and performance evaluation metrics.

In supposition, Figure 1 depicts a comprehensive image processing pipeline for medical image analysis. The process starts with data selection and pre-processing, which includes noise reduction and contrast enhancement. The image is then transformed with DWT and Haar wavelet, and key features are extracted for classification. This proposed system aims to improve diagnostic accuracy while also providing valuable information to medical professionals, making it an important tool in healthcare.

1.png

Figure 1. Block diagram of proposed system

The "BioSwarmNet" model addresses medical image classification challenges. This novel method integrates Fractional Order Differential Particle Swarm Optimisation (FODPSO) and Recurrent Neural Networks (RNNs). Unlike traditional methods that rely on manual hyperparameter tuning, FODPSO optimizes the architecture and parameters of RNNs using swarm intelligence and Darwinian principles. The incorporation of swarm intelligence and deep learning in BioSwarmNet provides a one-of-a-kind and integrated solution for medical image classification, which has the potential to improve diagnostic accuracy while reducing computational resources. The healthcare-focused application of this model suggests that it has the potential to aid medical professionals in disease detection and monitoring, making it a promising and innovative contribution to the field of medical image analysis.

The workflow shown in Figure 1 begins with a user-friendly dialogue that allows you to select an image file from the BRATS dataset [14] in formats such as jpg, bmp, gif, or png. Let f (x, y) represent the selected image file from the BRATS dataset, where x and y are pixel coordinates. Once chosen, the image is pre-processed, resizing to a standard 256×256 pixel dimension.

If f (x, y) is of size M×N, it is resized to a standard 256×256 pixel dimension as follows:

$f^{\prime}(x, y)=Resize(f(x, y),\ 256,256)$                  (1)

where, Resize (f(x, y), 256, 256) denotes the resizing operation.

Resizing images to a uniform size is a standard practice in image processing, especially in machine learning applications. This ensures that all images fed into the model have consistent dimensions, which is essential for many algorithms to function correctly. Uniformity in image size facilitates batch processing during neural network training and reduces computational load, which speeds up processing and improves efficiency.

Conversion to grayscale is meticulously performed in the case of colour images with three channels. In the case of color images with three channels (R, G, B), grayscale conversion is performed using the luminance method:

$f^{\prime \prime}(x, y)=0.2989 \cdot R+0.5870 \cdot G+0.1140 \cdot B$                  (2)

where, f′′(x, y) represents the grayscale image, and R, G, and B are the color channels at pixel (x, y).

Grayscale conversion simplifies the data by reducing it from three color channels to a single channel. This reduction lowers computational complexity, and shifts focus to intensity variations, which are often sufficient for feature detection in medical images.

Following that, the image goes on a process of noise reduction via the use of a Wiener filter. The Wiener filter operates in the frequency domain, and it can be represented mathematically as follows:

Let F(u, v) represent the two-dimensional Fourier transform of the pre-processed image f′′(x, y), where u and v are the frequency domain coordinates.

The power spectral density (PSD) of the noise in the image can be estimated as N(u, v).

The Wiener filter H(u, v) is computed as:

$H(u, v)=\frac{1}{H(u, v)}\left(\frac{|F(u, v)|^2}{|F(u, v)|^2+|N(u, v)|^2}\right)$                  (3)

where, H(u, v) is the Wiener filter in the frequency domain, ${\mid}F\left( u,v \right){{\mid }^{2}}$ is the squared magnitude of the Fourier transform of the pre-processed image, ${\mid}N\left( u,v \right){{\mid }^{2}}$ is the squared magnitude of the noise PSD.

Noise reduction is crucial in medical imaging where high-quality images are necessary for accurate diagnosis. The Wiener filter, a statistical approach, adjusts its effect based on the local image variance—performing minimal smoothing where variance is high, and more where it is low. This capability allows it to preserve essential edge details while reducing noise, a critical factor in medical image analysis for accurately delineating feature boundaries such as tumors.

The filtered image f′′′(x, y) is obtained by taking the inverse Fourier transform of the product of H(u, v) and F(u, v):

$f^{\prime \prime \prime}(x, y)=F^{-1}\{H(u, v) \cdot F(u, v)\}$                  (4)

where, F^-1 represents the inverse Fourier transform operation. This equation describes the application of a Wiener filter to reduce noise in the pre-processed image.

Subsequent that, an Adaptive Histogram Equalization step improves the image's quality, optimizing it for subsequent processes that culminate in the display of the filtered image [25].

AHE is applied to the pre-processed image f′′′(x, y) to improve its quality. It involves the following steps:

Compute the local histogram hi(k) for each pixel (x, y) in a small neighborhood window of size w×w:

$h i(k)=\sum_{p, q \in W i} \delta\left(f^{\prime \prime \prime}(p, q)-k\right)$                  (5)

where, Wi represents the set of pixels in the neighborhood window centered at pixel (x, y), and δ(⋅) is the Dirac delta function.

Compute the cumulative distribution function (CDF) for each local histogram Hi(k):

$H i(k)=\sum_{i=0}^k h i(j)$                  (6)

Calculate the transformation function Ti(k) for each local window:

$Ti(k)=\frac{(L-1)}{w^2} \sum_{j=0}^k \frac{\left(H_i(j)\right.}{H\left(w^2\right)-1}$                  (7)

Apply the transformation function Ti(k) to the pixel (x, y) in the neighborhood window:

$f^{\prime \prime \prime \prime}(x, y)=Ti\left(f^{\prime \prime \prime}(x, y)\right)$                  (8)

Repeat these steps for all pixels in the image to obtain the final enhanced image f′′′′(x, y).

AHE enhances image contrast, improving the visibility of features by better utilizing the dynamic range of intensities. This technique is especially valuable in medical imaging, where subtle contrast differences between tissues can be crucial for accurate diagnoses. AHE enhances these differences, aiding in the detection of features such as tumors or other anomalies.

Succeeding these preliminary steps, the pre-processed image is transformed using the Discrete Wavelet Transform (DWT) [25] and the Haar wavelet. The DWT yield consists of four coefficients denoted as ll, lh, hl, and hh, which are combined and displayed as a complete image [26].

Because of its simplicity and effectiveness, the Haar wavelet is commonly used for the DWT. The DWT splits the pre-processed image f′′′′(x, y) into four coefficient sets: ll (low-low), lh (low-high), hl (high-low), and hh (high-high). DWT is commonly represented as a series of convolution operations and down sampling.

First, define the Haar wavelet functions ψ(x) and ϕ(x) as:

$\psi(x)=\left\{\begin{array}{cc}\frac{1}{\sqrt{2}} & \text { if } 0 \leq x<1 / 2 \\ -\frac{1}{\sqrt{2}} & \text { if } 1 / 2 \leq x<1 \\ 0 &  Otherwise \end{array}\right.$                  (9)

$\phi(x)=\left\{\begin{array}{cc}

\frac{1}{\sqrt{2}} & \text { if } 0 \leq x<1 / 2 \\

\frac{1}{\sqrt{2}} & \text { if } 1 / 2 \leq x<1 \\

0 &  Otherwise

\end{array}\right.$                  (10)

Apply the DWT to f′′′′(x, y) using Haar wavelet to provide ll(x, y), lh(x, y), hl(x, y), and hh(x, y) which represent the low-low, low-high, high-low, and high-high coefficients, respectively [27, 28].

Combine these coefficients to reconstruct the image:

$\begin{gathered}f^{\prime \prime \prime \prime \prime}(x, y)=  Combine (\operatorname{ll}(x, y), \operatorname{lh}(x, y), \operatorname{hl}(x, y), \operatorname{hh}(x, y))\end{gathered}$                  (11)

The function Combine (⋅) represents the inverse DWT operation, which combines the coefficients to reconstruct the image f′′′′′(x, y).

The result f′′′′′(x, y) is a composite image formed by combining the four sets of DWT coefficients, ready for further processing or display. In terms of specifics, the DWT's ll (low-low) coefficients hold the key to feature extraction [29, 30].

The Grey Level Co-occurrence Matrix (GLCM) is computed using the coefficients obtained from the Discrete Wavelet Transform (DWT) as follows:

Let ll(x, y) represent the low-low coefficients obtained from the DWT. Define the GLCM for a specific distance d and direction θ as P_d,θ (i, j), where i and j are gray levels.

Compute the GLCM for the given distance d and direction θ by counting the occurrences of pairs of gray values in ll(x, y) that meet the specified conditions:

${{P}_{d,\theta }}\left( i,j \right)=\mathop{\sum }_{x}\mathop{\sum }_{y\text{ }\!\!~\!\!\text{ }}\left\{ \begin{matrix}   1,~~if~ll\left( x,y \right)=i~and~ll\left( x+\Delta x,y+\Delta y \right)=jfor~the~specified~\Delta x,\Delta y,~and~\theta   \\0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~elsewhere  \\\end{matrix} \right.\text{ }\!\!~\!\!\text{  }\!\!~\!\!\text{  }\!\!~\!\!\text{  }\!\!~\!\!\text{  }\!\!~\!\!\text{ }$                  (12)

Normalize the GLCM by dividing each element by the total number of valid pairs:

$P_{d, \theta}(i, j)=\frac{P_{d, \theta}(i, j)}{\sum_i \sum_j P_{d, \theta}(i, j)}$                  (13)

Once the GLCMs are computed, a set of statistical properties, often referred to as "Query Features," can be extracted. These features include:

$\text{Energy}E=\sum i \sum j P_{d \theta}(i, j)^2$                  (14)

$Contrast\ (C): C=\sum i \sum j(i-j)^2 P_{d, \theta}(i, j)$                  (15)

$Correlation\ (CO): C O=\frac{\left[\sum i \sum j(i-\mu)(j-v) P d, \theta(i, j)\right]}{\sigma i \sigma j}$                  (16)

where, μ and ν are the means of i and j respectively, and σi and σj are their standard deviations.

$Homogeneity\ (H): H=\sum_i \sum_j \frac{P_{d, \theta}(i, j)}{1+|i-j|}$                  (17)

$\begin{gathered} Entropy\ (E N T): ENT=-\sum i \sum j P_{d \theta}(i, j) \log _2\left(P_{d, \theta}(i, j)+\epsilon\right)\end{gathered}$                  (18)

where, ϵ is a small positive constant to avoid the logarithm of zero.

This group of features is then assigned to the proposed BioSwarmNet model. A pretrained neural network model takes centre stage in the classification process. This neural network categorises the image as normal, abnormal tumour, benign tumour, or malignant tumour.

Tailored actions take place based on the outcome of this classification (C). If C is 1 or 3, a user-friendly message dialogue appears, explaining whether the image represents a normal case for the relevant organ. If C is 2 or 4, a warning message appears on the screen, while tumour localization and area calculations are meticulously carried out behind the scenes. If a tumour is detected, the original image begins a segmentation process.

The boundaries of the segmented tumour are accurately delineated during this segmentation process, and the original image is adorned with these distinct contours, making the tumor's presence visually apparent. The tumour region is effectively isolated and displayed independently, and its size is precisely calculated.

Finally, the workflow delves into performance evaluation, calculating a number of key performance metrics such as True Positive (TP), True Negative (TN), False Positive (FP), False Negative (FN), Sensitivity, Specificity, and Accuracy. These metrics are extremely useful in determining the system's efficacy and accuracy.

3.1 Proposed model

To address the challenges of medical image classification, the BioSwarmNet model as provided in Figure 2 combines two distinct paradigms, Fractional Order Differential Particle Swarm Optimisation (FODPSO) and Recurrent Neural Networks (RNN).

The term "Bio" in BioSwarmNet refers to the FODPSO bio-inspired optimization technique. FODPSO is inspired by Darwinian principles and swarm intelligence, making it an efficient optimization method for complex problems. PSO and other swarm intelligence algorithms are inspired by the collective behavior of social organisms. FODPSO, a PSO variant, applies a swarm-based optimization strategy to medical image classification. The word "Net" in the title emphasizes the incorporation of neural networks, specifically RNNs. The capabilities of neural networks in feature learning and classification tasks are well known.

2.png

Figure 2. Proposed BioSwarmNet model architecture

By combining Fractional Order Differential Particle Swarm Optimisation (FODPSO) and Recurrent Neural Networks (RNNs), BioSwarmNet presents a novel and integrated approach to medical image classification. Its uniqueness stems from FODPSO, a bio-inspired optimisation strategy that optimises RNN architectures and parameters, as opposed to traditional methods that frequently rely on manual hyperparameter tuning. This combination of swarm intelligence and deep learning addresses an important need in healthcare, with the potential to improve diagnostic accuracy while optimising computational resources. Its healthcare-focused application highlights its potential to assist medical professionals in disease detection and monitoring, making it a promising and innovative contribution to the field.

3.2. Algorithm of the proposed BioSwarmNet model

BioSwarmNet provides a holistic approach by seamlessly combining optimisation and feature learning, potentially improving efficiency and accuracy in medical image analysis and its stepwise algorithm is provided below.

Finally, this research paper provides a thorough examination of an advanced medical image analysis system, which is supported by a novel and innovative model known as "BioSwarmNet." The proposed methodology includes an image processing pipeline that has been meticulously designed to ensure data uniformity and quality. The system starts with user interaction to select a dataset and then moves on to pre-processing steps such as resizing, grayscale conversion, noise reduction, and feature extraction using the Discrete Wavelet Transform (DWT) and Grey Level Co-occurrence Matrix (GLCM). This comprehensive workflow concludes with the classification of medical images, allowing for tailored actions based on the results, such as tumour localization and performance evaluation metrics. Notably, the proposed system has achieved remarkable results in key metrics, including 99.12% accuracy, 98.62% sensitivity, and 99.86% specificity. Achieving high accuracy, sensitivity, and specificity with the "BioSwarmNet" model translates into substantial real-world clinical benefits. Enhanced accuracy ensures reliable diagnoses, reducing unnecessary treatments and improving patient outcomes. High sensitivity aids in early disease detection, increasing treatment effectiveness, while high specificity minimizes false positives, preventing undue patient stress and reducing wasteful healthcare spending. Collectively, these improvements streamline healthcare workflows and elevate the standard of patient care.

The revolutionary "BioSwarmNet" model is a hybrid of Fractional Order Differential Particle Swarm Optimisation (FODPSO) and Recurrent Neural Networks (RNNs). By combining the power of swarm intelligence and deep learning for automated medical image classification, this model has the potential to revolutionise the field of medical image analysis. The integration of FODPSO for automated hyperparameter tuning with RNNs in BioSwarmNet demonstrates its ability to improve diagnostic accuracy in a healthcare-focused application. Furthermore, the study made use of the well-known and preferred BRATS dataset, which added credibility and relevance to the research. Adopting a multi-disciplinary approach is crucial for the further development of the "BioSwarmNet" model. Collaboration between computer scientists, radiologists, and other healthcare professionals will enhance the model's accuracy and clinical applicability. Such teamwork ensures algorithmic improvements are clinically relevant and aligned with healthcare standards, leading to a robust, universally effective diagnostic tool.