TLDR: A new deep learning framework for diagnosing ischemic and hemorrhagic brain strokes from CT images has been developed. The study utilized pre-trained models for feature extraction, combined with feature engineering techniques and traditional machine learning classifiers. The most effective combination, MobileNetV2 with Linear Discriminant Analysis (LDA) and a Support Vector Classifier (SVC), achieved 97.93% accuracy, demonstrating a highly accurate and computationally efficient method for stroke detection.
Brain stroke remains a critical global health challenge, being a leading cause of death and long-term disability. Early and accurate diagnosis is crucial for effective treatment and improved patient outcomes. Traditionally, medical professionals rely on clinical evaluations and imaging techniques like Computed Tomography (CT) scans for diagnosis. However, the advent of machine learning (ML) and deep learning (DL) offers promising new avenues to enhance these diagnostic processes.
A recent study introduces a novel deep learning framework designed to improve the diagnosis of ischemic and hemorrhagic brain strokes using CT images. The research highlights that while many existing stroke classification methods focus on single CT slices, a more comprehensive approach is needed to assist radiologists in identifying critical areas from full CT volumes.
The core of this new framework involves leveraging pre-trained deep learning models for feature extraction from CT scan images. The researchers utilized several well-known models, including DenseNet201, InceptionV3, MobileNetV2, ResNet50, and Xception. These models are adept at identifying intricate patterns and features within medical images. To further refine the extracted features and boost performance, the study incorporated advanced feature engineering techniques such as Bacterial Foraging Optimization (BFO), Principal Component Analysis (PCA), and Linear Discriminant Analysis (LDA).
Following feature extraction and optimization, these refined features are then fed into various traditional machine learning algorithms for classification. The classifiers tested included Support Vector Classifier (SVC), Random Forest (RF), Extreme Gradient Boosting (XGB), Decision Tree (DT), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Gaussian Naive Bayes (GNB).
The researchers meticulously prepared a robust dataset for this study by merging two publicly available datasets and supplementing them with additional normal brain images from a local clinic in Bangladesh. This curated dataset, comprising 3819 CT scan images, was categorized into Normal, Ischemic, and Hemorrhagic classes, ensuring a comprehensive basis for training and evaluation.
Through extensive experimentation, the study identified a standout combination: MobileNetV2 for feature extraction, Linear Discriminant Analysis (LDA) for feature optimization, and Support Vector Classifier (SVC) for classification. This specific combination achieved an impressive classification accuracy of 97.93%. This result significantly outperformed other tested model-optimizer-classifier combinations, demonstrating the effectiveness of integrating lightweight pre-trained models with robust optimization and classification techniques for brain stroke diagnosis.
The findings suggest that this feature extraction-based approach offers a more efficient and scalable solution compared to computationally intensive end-to-end deep learning models, making it a practical choice for real-world clinical applications where computational resources might be limited. The model’s ability to balance high classification performance with computational efficiency is a key advantage. For more details, you can refer to the full research paper here.
Also Read:
- Advanced AI System Achieves High Accuracy in COVID-19 Diagnosis from Lung CT Scans
- LangMamba: Enhancing Low-Dose CT Denoising with Vision-Language Models
While the proposed model shows strong performance, the authors acknowledge certain limitations, such as potential limitations in adaptability to highly diverse datasets due to the feature extraction approach rather than fine-tuning. Future research could explore self-supervised learning, transformer-based architectures, and the incorporation of multi-modal data sources to further enhance diagnostic accuracy and generalizability.
