System for early prediction of brain tumors using a deep machine learning model

Abstract

A deep learning-based system for the detection of brain tumors, consisting of: i. a preprocessing module for normalizing, segmenting and enhancing MRI images; ii. a CNN architecture optimized for spatial feature extraction and classification; iii. a training pipeline with data set partitioning and hyperparameter tuning; iv. a visualization module for explainability and model interpretability.

Description

AREA OF INVENTION

[0001] This invention relates to a system for the early prediction of brain tumors using a deep machine learning model. BACKGROUND OF THE INVENTION

[0002] Brain tumors are considered a deadly disease, affecting more than 300,000 people worldwide each year. Rapid diagnosis is essential for successful treatment, improving both therapy outcomes and patient survival rates. Current diagnostic approaches rely on manual feature selection and conventional techniques, making them prone to errors and requiring lengthy processing times. Artificial intelligence and deep learning technologies have recently advanced to the point where they can automate tumor detection in MRI scans while simultaneously improving detection accuracy.

[0003] Deep learning models, especially convolutional neural networks (CNNs), are frequently used by researchers to analyze MRI images. The reason CNNs are superior to conventional methods lies in their ability to automatically extract hierarchical features directly from the image data. However, all existing implementations suffer from the following limitations. Tumor detection in MRI scans was performed using pre-trained CNNs fine-tuned with the Manta Ray Foraging Optimization algorithm. However, because their method also relied on large pre-trained models, they achieved high precision but at a significant computational cost, making real-time applications impossible [1]. MobileNetV3 (and other transfer learning frameworks) have been used to classify tumors using MRI.An accuracy of 99.75% was achieved, but could not be generalized to external data sets due to over-fitting to the input database [2].

[0004] Brain tumor segmentation was performed using transfer learning and YOLO (You Only Look Once), two advanced object recognition algorithms. In another study, the YOLOv5 and YOLOv7 algorithms were used for brain tumor detection and classification. Due to their high sensitivity and specificity, the models required significant computational resources for training and inference [3]. An improved ResNet architecture was employed to address the vanishing gradient problem in deep neural networks. While the proposed method improved segmentation accuracy, it proved computationally intensive for use in resource-constrained environments, such as the requested planetary services [4]. Certain preprocessing techniques were also explored to enhance the image quality of MRI images and increase model robustness.For example, improved use of MRI scans with more filters and data extension for the identification of pituitary tumors was investigated. While the proposed technique led to improved accuracy, it was computationally intensive [5]. In another study, InvNets were used for medical image processing tasks. Interestingly, InvNets achieved higher accuracy and required fewer parameters than CNNs, but their extreme fine-tuning requirements made them unsuitable for practical application in clinical workflows [6].

[0005] Several studies have been conducted on the multi-class classification of tumor types. An architecture based on EfficientNet-B4 was proposed, achieving an accuracy of 99.33%. However, this approach still required multiple layers and consumed significant computing power, preventing its use in real-time diagnostic systems [7]. Similarly, GLCM (Gray Level Co-occurrence Matrix) was used to extract the feature and improve the segmentation accuracy in CNN models. While this method proved promising, it required a high preprocessing overhead [8]. US20210000567A1: An invention relating to “Deep Learning System for Medical Image Classification” discloses a generic CNN for tumor detection, which, however, lacks specificity in preprocessing and does not take into account the variability of the data set. EP3451234B1: Another invention on the topic of "Method for segmenting brain tumors using transfer learning" uses pre-trained models such as ResNet, but exhibits high computational costs and overfitting with small datasets. CN113963134A: A similar invention for "Medical image analysis using hybrid neural networks" combines CNNs and RNNs, but does not optimize preprocessing or validation for clinical reliability. Summary of the invention

[0006] Previous research in this area has several important limitations. Data imbalances are a major problem in many studies, as class imbalances in the datasets have not been accounted for, leading to biased predictions. Existing models also exhibit insufficient generalization, as they are difficult to adapt to changes in MRI acquisition protocols. Furthermore, the high computational cost of training complex architectures and pre-trained models is a significant obstacle. Additionally, robust preprocessing techniques capable of effectively handling noise and artifacts in MRI scans are lacking, reducing the overall reliability of MRI models.

[0007] This invention provides a CNN-based system that overcomes the limitations of the prior art by: 1. Use of robust data preprocessing techniques to normalize and segment MRI images for accurate analysis. 2. Use of data augmentation methods to improve model generalization and reduce overfitting. 3. Integration of an optimized CNN architecture to improve feature extraction and classification accuracy. 4. Achieving high-performance metrics, including 99% accuracy, sensitivity, and precision, validated using balanced datasets. 5. Support for clinical integration with low computational overhead and compatibility with resource-constrained environments. DETAILED DESCRIPTION OF THE INVENTION

[0008] The invention comprises a comprehensive, deep learning-based diagnostic system for the detection of brain tumors with advanced preprocessing, optimized CNN architecture and rigorous evaluation methods.

[0009] The dataset comprises 4,600 MRI images of the axial, coronal, and sagittal planes. Preprocessing ensures consistency of the input dimensions, with all images scaled to 256 × 256 pixels. Intensity normalization converts the pixel values ​​to a uniform scale, thus reducing acquisition-related variability. Segmentation isolates brain tissue and removes irrelevant regions to improve the model's focus on tumor-specific features.

[0010] The CNN architecture follows a structured hierarchical approach. The input layer processes pre-processed MRI scans and standardizes them before feature extraction. Convolutional layers apply multiple filters to capture spatial patterns and tumor-specific features. Pooling layers reduce spatial dimensions while preserving important information, thus optimizing computational performance. Dense layers further process extracted features and classify images as healthy or tumor-infected.

[0011] The model is trained on a 70:20:10 dataset for training, validation, and testing. Hyperparameter tuning techniques, including early stopping and learning rate adjustment, optimize model convergence and prevent overfitting. Performance evaluation includes receiver operating characteristic (ROC) curves, confusion matrices, and F1 score metrics, ensuring a comprehensive assessment of classification reliability.

[0012] A key aspect of the invention is its low computational effort compared to alternative deep learning models such as YOLOv7 and EfficientNet. The proposed architecture reduces inference time while maintaining high recognition accuracy and is therefore suitable for real-time applications in clinical settings.

[0013] Furthermore, the invention integrates explainability functions and utilizes activation heatmaps and convolutional feature maps to highlight tumor-specific regions in MRI scans. This increases transparency and supports clinical decision-making through visual evidence of tumor classification.

[0014] The system is designed for seamless clinical integration and features an intuitive user interface for medical professionals. The model's adaptability allows for use in hospitals, diagnostic centers, and telemedicine platforms, bridging the gap between research advances and practical healthcare applications. 1. Data set:

[0015] The system processes data from 4,600 MRI images, including samples from both healthy and tumor patients, in axial, coronal, and sagittal perspectives. The system applies preprocessing steps that set the images to a resolution of 256 × 256 pixels and simultaneously normalize the pixel intensity values ​​to a scale of 0 to 1. 2. Preprocessing:

[0016] The preprocessing includes: • Segmentation to isolate brain tissue. • Data enhancement techniques to avoid overfitting, including mirroring, rotating, and brightness modulation to enlarge training data and prevent overfitting. 3. CNN architecture:

[0017] The CNN model consists of the following components • Input layer: The preprocessed MRI images are entered into this layer. The method removes non-brain regions and then normalizes the pixel intensity to reduce acquisition-related variability. • Convolution layers: These layers are used to extract spatial features using various filters. • Pooling levels: The pooling levels reduce spatial information without losing important properties of the data. • Dense layers: Dense layers perform pattern recognition and classification tasks. • Output level: This level assigns a binary classification (healthy / tumor). 4. Training and Validation

[0018] The dataset is divided into three parts for training, validation, and test sets in a ratio of 70:20:10. To avoid the problem of overfitting, techniques for early stopping and model checking were used during training. 5. Key Performance Indicators

[0019] The model achieves an accuracy of 99%, a precision of 99%, a sensitivity (recall) of 99%, a specificity with high detection rates for healthy cases, and an F1 score by harmonizing precision and recall for robust performance evaluation. 6. Visualization

[0020] The system's detection capabilities for tumor-specific areas are verified by convolutional layer feature maps in combination with activation heatmaps, preparing the platform for clinically explainable use. References 1) Aljohani, M., Bahgat, WM, Balaha, HM, AbdulAzeem, Y., El-Abd, M., Badawy, M., & Elhosseini, MA (2024) An automated metaheuristic-optimized approach for the diagnosis and classification of brain tumors based on a convolutional neural network. Results in Engineering, 23(102459), 102459. DOI: 10.1016 / i.rineng.2024.102459 2) Mathivanan, SK, Sonaimuthu, S., Murugesan, S., Rajadurai, H., Shivahare, BD, & Shah, MA (2024) Application of deep learning and transfer learning for accurate detection of brain tumors. Scientific Reports, 14(1), 7232. DOI: 10.1038 / s41598-024-57970-7 3) Almufareh, MF, Imran, M., Khan, A., Humayun, M., & Asim, M. (2024) Automated segmentation and classification of brain tumors in MRI using YOLO-based deep learning. IEEE Access, 12, 16189-16207. DOI: 10.1109 / access.2024.3359418 4) Aggarwal, M., Tiwari, AK, Sarathi, MP, & Bijalwan, A. (2023) Early detection and segmentation of brain tumors using a deep neural network. BMC Medical Informatics and Decision Making, 23(1), 78. DOI: 10.1186 / s12911-023-02174-8 5) Abdusalomov, AB, Mukhiddinov, M., & Whangbo, TK (2023) Brain tumor detection based on deep learning approaches and magnetic resonance imaging. Cancers, 15(16), 4172. DOI: 10.3390 / cancers15164172 6) Asiri, AA, Shaf, A., Ali, T., Zafar, M., Pasha, MA, Irfan, M., Alqahtani, S., Alghamdi, AJ, Alghamdi, AH, Alshamrani, AFA, Aleylyani, M., & Alamri, S. (2023) Improving brain tumor diagnosis: Transition from the convolutional neural network to the involutional neural network. IEEE Access, 11, 123080-123095. DOI: 10.1109 / access.2023.3326421 7) Preetha, R., Priyadarsini, MJP, & Nisha, JS (2024) Automatic brain tumor detection using magnetic resonance imaging with the finely tuned EfficientNet-B4 convolutional neural network. IEEE Access, 12, 112181-112195. DOI: 10.1109 / access.2024.3442979 8) Sharma, M., & Miglani, N. (2020) Automated brain tumor segmentation in MRI images using deep learning: Overview, challenges and future. In Studies in Big Data (pp. 347-383). Springer International Publishing. 9) Patent US20210000567A1 10) Patent EP3451234B1 11) Patent CN113963134A QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 20210000567A1 [0005, 0020] EP 3451234B1 [0005, 0020] CN 113963134A [0005, 0020] Cited non-patent literature

[0000] Brain tumors in MRI using YOLO-based deep learning. IEEE Access, 12, 16189-16207. DOI: 10.1109 / access.2024.3359418

[0020] Transition from Convolutional Neural Network to Involutional Neural Network. IEEE Access, 11, 123080-123095. DOI: 10.1109 / access.2023.3326421

[0020] Preetha, R., Priyadarsini, MJP, & Nisha, JS (2024) Automatic brain tumor detection from magnetic resonance images using the finely tuned EfficientNet-B4 Convolutional Neural Network. IEEE Access, 12, 112181-112195. DOI: 10.1109 / access.2024.3442979

[0020] Sharma, M., & Miglani, N. (2020) Automated brain tumor segmentation in MRI images using deep learning: Overview, challenges and future. In Studies in Big Data (pp. 347-383). Springer International Publishing

[0020]

Claims

[1] A deep learning-based system for the detection of brain tumors, consisting of: i. a preprocessing module for normalizing, segmenting and enhancing MRI images; ii. a CNN architecture optimized for spatial feature extraction and classification; iii. a training pipeline with data set partitioning and hyperparameter tuning; iv. a visualization module for explainability and model interpretability. [2] System according to claim 1, wherein the preprocessing module standardizes MRI images to 256 × 256 pixels and normalizes pixel intensity values. [3] System according to claim 1, wherein the CNN architecture consists of convolutional layers, pooling layers and fully connected dense layers. [4] System according to claim 1, wherein data extension techniques, including mirroring, rotation and brightness modulation, improve model generalization.