ENHANCING DIAGNOSTIC PRECISION WITH AI: BRAIN TUMOR DETECTION IN HEALTHCARE

Abstract

Brain tumour classification using DL has shown significant progress in recent years, with promising results in terms of accuracy and automation. However, addressing challenges related to data availability, interpretability, and clinical integration will be crucial for successful adoption in real-world healthcare settings. Overall, our findings show that good performance on the Brain tumour dataset is 96.5% accuracy when compared to other existing works. Further, we improved accuracy by applying the Adam optimization technique which obtained 98% accuracy on the classification of brain tumour using CNN. Further research and advancements in deep learning techniques are expected to improve the performance and utility of brain tumour classification systems, ultimately benefiting patients and medical professionals alike. One of the major Constraints of this study is the limited dataset size, the availability of large, annotated datasets for brain tumour classification remains a challenge, which can hinder the generalization and robustness of DL models. DL models are often considered black boxes, making it challenging to interpret their decisions and establish trust in the clinical setting.

Specification

Description:FIELD OF INVENTION:
[0001] The main concept of this research is "BRAIN TUMOR DETECTION IN HEALTHCARE by ENHANCING DIAGNOSTIC PRECISION WITH AI".

BACKGROUND:
[0002] The Brain tumours are a complex and heterogeneous disease group requiring accurate classification for effective treatment planning. Deep learning (DL) techniques, specifically CNNs, have gained attention in recent years due to their ability to automatically learn and extract high-level features from medical images. Brain tumours are masses that develop when the brain's regulatory processes are destroyed, allowing for the uncontrolled multiplication of brain cells. Skull tumours, if allowed to progress, can impose harmful pressure on the brain and compromise overall health.

[0003] Brain tumours are a leading cause of cancer-related mortality in the United States, making early detection and categorization a critical area of research in the field of medical imaging. There are numerous approaches to categorise malignant brain tumours. Brain tumours, for instance, are commonly divided into two categories: benign and malignant. Benign brain tumours typically form in the skull itself, away from the brain itself.

SUMMARY:
[0004] A brain tumour classifier using DL techniques is an advanced approach that can automatically classify brain tumours from medical images such as MRI scans. Deep learning techniques, such as CNNs, RNNs, and DBNs, can automatically learn complex patterns and features from medical images and thus are ideal for image classification tasks. The classifier works by first training a deep learning model on a large dataset of labelled brain MRI scans. During training, the model learns to automatically extract important features and patterns from the MRI scans and use them to accurately classify the images into different categories, such as tumour or non-tumour.

[0005] Once the deep learning model is trained, it can be used to classify new, unseen brain MRI scans into tumour or non-tumour categories. The classifier takes the MRI scan as input and applies a series of deep learning techniques to extract and learn the most relevant features of the image. The model then passes these features through one or more fully connected layers to make the final classification decision. Compared to traditional ML approaches, DL-based classifiers have demonstrated superior performance in classifying brain tumours. They can automatically learn complex features and patterns from medical images, which may not be easily detectable by human observers. This can help in early and accurate detection of brain tumours.

[0006] Overall, a brain tumour classifier using DL techniques has the potential to significantly raise the accuracy and efficiency of brain tumour diagnosis, which can ultimately lead to better patient care and outcomes.A brain tumour classifier using CNNs classifies brain tumours from medical images such as MRI scans. The classifier works by first training a CNN model on a dataset of labelled brain MRI scans. The CNN model consists of multiple layers of filters that are designed to detect features at different levels of abstraction in the input image. During training, the CNN learns to extract these features and use them to classify the input images into different categories (i.e. tumour or non-tumour).

[0007] Once the CNN model has been trained, it may be used to classify new brain MRI scans into the tumour or non-tumour category depending on whether or not they have previously been seen. The MRI scan is utilised as the input for the classifier, which then proceeds to extract features from the image using a combination of convolutional and pooling layers. After that, these traits are sent via one or more layers that are completely connected, which ultimately decides how they should be classified.

[0008] Compared to traditional ML approaches, dl-based classifiers using CNNs have demonstrated superior performance in classifying brain tumours. They can automatically learn complex features and patterns from medical images, which may not be apparent to human observers.

[0009] This research aims to discern the optimal CNN model for brain tumour classification using publicly available datasets and MRI images, employing CNN techniques wherein a vast majority of hyper-parameters are automatically fine-tuned via the Adam optimizer. To attain this objective, extensive utilization of CNN methodologies will be a cornerstone of this study. The ensuing structure of this report is as follows: The subsequent section will juxtapose our findings with analogous conclusions drawn by other studies. Section 3 will provide a comprehensive exposition of the proposed CNN models. Section 4 will delve into the experimental results and offer a meticulous comparison with state-of-the-art alternatives. Finally, Section 5 will mark the concluding segment of this paper.

DESCRIPTION OF FIGURES AND TABLES:
[0010] Table I provides a general comparison of techniques for brain tumour detection, highlighting their advantages, disadvantages, and common applications.

[0011] Table ii. Research Investigations carried out in the realm of Brain Tumor Detection.

[0012] Fig. 1. A) Normal Brain MRI images.

[0013] Fig. 1. B) Abnormal Brain Images.

[0014] Fig.2. Block Diagram of Proposed System Model.

[0015] Fig.3. Confusion Matrix

[0016]. Table iii. THE PERFORMANCE COMPARISON WITH OTHER EXISTING FINDINGS.

[0017] Fig.4. Evaluation of models in terms of Accuracy

[0018] Fig 5 a) Pre-processed images of Normal MRI b) Tumour MRI c) Pre-processed Normal MRI d) Tumour MRI e) Pre-processed Normal MRI f) Tumour MRI

[0019] Fig. 6. Comparison of Accuracy Metric.

[0020] Fig 7: Loss graphs for training and validation.

[0021] Fig 8: Accuracy graphs for training and validation.

[0022] Fig. 9. Evaluation of performance metrics for Proposed and Existing works.

[0023] Fig. 10. Classification outcome produced by the CNN model.
[0024] Fig. 11. Computation time for Proposed Vs. Existing approaches.

[0025] Table IV. The Performance of all the models.

[0026] Table V. Comparing the proposed work with previous.


METERIALS AND METHODS
[0027] Dataset Details:In this, we present a machine learning-based classification model that utilizes patient MRI scans as input to determine the presence or absence of a brain tumour. To obtain the MRI scans, we utilized the freely available dataset from Kaggle, which consists of a total of 253 brain MRI scans. Among these scans, 155 images are classified as "yes" indicating the presence of brain tumours, while 98 images depict normal brain tissue and are classified as "no". In Figure 1, it is illustrated that 155 samples, accounting for 61% of the dataset, are considered exemplary, while the remaining 98 samples, representing 39% of the data, are deemed problematic.

[0028]. Convolutional Neural Networks (CNNs): A CNN is a specialized DL model renowned for its efficacy in tasks involving image and video analysis. CNNs find extensive application in computer vision domains such as object recognition, image classification, image segmentation, and beyond. The proposed model's comprehensive architecture is depicted in Figure 2.

[0029] Convolutional Layers: To put it simply, a CNN is only as good as its convolutional layers. Feature maps are generated by conducting element-wise multiplication and summing operations on the input data using a collection of learnable filters (sometimes called kernels). The data's local patterns and geographical linkages are shown by these filters. The training procedure imparts knowledge of the filters, which the network then uses to discover pertinent features on its own. Using equation (1), we can get the pre-nonlinearity input to some unit xlij in our layer by adding the contributions (weighted by the filter components) of the cells in the preceding layer.

x_ij^l=∑_(a=0)^(m-1)▒∑_(b=0)^(m-1)▒w_(ab y_(i+a)(j+b)^(l-1) ) ………………………………….(1)

[0030] Pooling Layers: Pooling layers are utilized to perform down sampling on the feature maps generated by the Conv layers. They maintain the most crucial data from the feature maps despite drastically decreasing their spatial dimensions. With the aid of pooling, the representation can be made more condensed and robust against subtle spatial shifts.

[0031] The max-pooling layers are non-learning and quite straightforward. They just select a kk range and return the maximum value inside that range. The max function reduces each kxk block to a single value, therefore if the input layer is a NxN layer, the resulting layer will be a N/kxN/k layer.

[0032] Activation Functions: By introducing non-linearities in the form of activation functions, CNN is able to simulate intricate connections between the input data and the features it has learned. Rectified Linear Unit (ReLU), sigmoid, and tanh are only some of the most common activation functions utilized in CNNs. Because of its ease of use and ability to solve the vanishing gradient problem, the ReLU activation function (expressed by equation (2)) is the most popular choice for this purpose.

[0033] Fully Connected Layers: The CNN's final stages typically consist of fully connected layers. They process the final classification or regression using the retrieved high-level characteristics from earlier layers.

[0034] Loss Function: It gauges the disparity between the CNN's predicted output and the actual labels, quantifying errors and steering the learning procedure. Typical loss functions for classification involve categorical cross-entropy and SoftMax, while mean squared error (MSE) is prevalent for regression tasks. Our approach involves passing the SoftMax layer's output, denoted as Oi = σi(Z), to the Loss layer, necessitating the initial computation of the top layer, equation (3) represents the SoftMax loss function.
∂l(y,o)/〖∂o〗_i = - δ_iy/o_y ………………(3).
As, we can utilize the chain rule:
〖∂o〗_i/〖∂z〗_k = (δ_ik e^(z_i ) (∑_(j=1)^m▒e^(z_i ) )-e^(z_i ) e^(z_k ))/(∑_(j=1)^m▒e^(z_i ) )^2 …………………… (4).
= δ_ik o_k-o_i o_k

[0034] Training: CNNs are trained using large labelled datasets through a process called backpropagation. The weights of the network are adjusted iteratively to minimize the loss function using optimization algorithms. The training process involves feeding the input data through the network, calculating the loss, propagating the error backwards, and updating the weights.

[0035] Adam Optimizer: Layers that are completely connected are often inserted towards the very end of the CNN. They are given the high-level features that were retrieved by the layers that came before them. Adam, which stands for adaptive moment estimation, is an optimisation procedure that is frequently utilised in deep learning to update the parameters of neural networks. It brings together concepts from momentum-based approaches like RMSprop and adaptive learning rate methods like AdaGrad. Adam can use the following equation as a final regression and classification analysis formula.
m_t=β_1 m_(t-1)+(-β_1 )[δL/〖δw〗_t ] ν_t=β_2 ν_(t-1)+(1-β_2 ) [δL/〖δw〗_t ]^2………………(5)

[0036] The algorithm works as follows: Set initial values for the neural network's parameters (weights and biases). Determine the parameter-depending gradient of the loss function. Initialize the first and second moments (vectors) as zero. These moments are estimates of the mean and uncentered variance of the gradients. Update the first moment (momentum) by exponentially decaying the previous first moment and adding the current gradient. Update the second moment (uncentered variance) by exponentially decaying the previous second moment and adding the squared current gradient. Compute bias-corrected first and second moment estimates to account for their initialization at zero. Update the parameters by moving in the direction of the gradients scaled by the adaptive learning rate, which is computed as the ratio of the bias-corrected first moment and the square root of the bias-corrected second moment, with a small constant added to avoid division by zero. Reiterate steps 2 to 7 for a designated no. of iterations or until convergence is attained.

[0037] PERFORMANCE EVALUATION: Evaluation of classification performance in image classification studies is of utmost importance to establish the scientific validity and credibility of the study results. Without a thorough performance evaluation, the classification study would remain incomplete and lack academic rigour.

[0038] In the realm of image classification studies, several well-established performance evaluation metrics have been widely adopted as standard benchmarks. These metrics, including accuracy, specificity, sensitivity, and precision, have stood the test of time and are consistently utilized in similar studies. By employing these recognized metrics, researchers can effectively compute the accuracy and reliability of the classification process, ensuring robust evaluation in this paper.

[0039] Furthermore, the evaluation of the models in this research is conducted using the AUC of the ROC curve. The AUC of the ROC curve is a well-established metric that quantifies the discriminatory capability of the model and its capacity to differentiate between various classes. Fig 3 displays the Confusion matrix, where TP, TN, FP, and FN respectively.

[0040] As part of our brain tumour classification using DL, we developed a binary classifier that can identify brain tumours by analysing MRI scans. We constructed our classifier with CNN and Adam optimizer, and as a result, an accuracy level of 98% was attained, demonstrating the comprehensive performance visualization of our model. The loss function selected for this purpose is categorical cross-entropy, which is crucial in deep learning model development. It serves to quantify the disparity between predicted and actual outputs, enabling neural networks to adapt their weights and enhance accuracy. Categorical cross-entropy specifically applies when categorizing an image into multiple classes. In neural networks, optimizers are primarily employed to adjust weights and expedite learning, ultimately minimizing losses. Adam stands out as the most widely utilized choice among the various optimisers available. , Claims:Claims:
1. A method of detecting brain tumour in healthcare by enhancing diagnostic precision, comprising:
a kaggle approach;
the kaggle consists of brain MRI scans;
the MRI scans are pre-processed and features are extrcated;
the extracted features are provided to convolution neural network;
whereby high-level features that were retrieved by the convolution neural network are optimized to update the parameters of neural networks;
the optimized features are predicted using predictive models;
wherein brain MRI scans images are classified as "yes" indicating the presence of brain tumours; and
images depict normal brain tissue and are classified as "no".

Documents