Automated brain tumor detection in MRI images using CNN and ResNet architectures

Technical Summary: Automated Brain Tumor Detection in MRI Images Using CNN and ResNet Architectures

Problem Statement
The early and accurate diagnosis of brain tumors is critical for determining treatment modalities and improving patient survival rates. While Magnetic Resonance Imaging (MRI) is the standard modality for visualizing brain structures due to its high soft-tissue contrast, the manual interpretation of these scans by radiologists remains labor-intensive, subjective, and prone to inter-observer variability. Traditional machine learning approaches relying on handcrafted features often fail to generalize across different datasets and imaging settings. Furthermore, the increasing volume of medical imaging data places a significant strain on clinical practitioners, raising the risk of missed or delayed diagnoses. There is an urgent need for automated, reliable, and cost-effective deep learning systems capable of distinguishing tumor-positive from tumor-negative MRI scans with high accuracy.

Methodology
The proposed framework utilizes a deep learning approach based on Convolutional Neural Networks (CNNs) and Residual Networks (ResNets) to automate brain tumor detection. The study employs transfer learning, leveraging pre-trained models to address the challenge of limited labeled medical data.

• Dataset: The system was evaluated on a dataset of 3,929 brain MRI images, classified by clinical experts into tumor and non-tumor categories. The data was split into 85% for training and 15% for testing.
• Preprocessing: Images were resized to 224 × 224 pixels to match the input requirements of the pre-trained models. Pixel intensities were normalized to improve convergence. Data augmentation techniques, specifically random horizontal flipping and rotation, were applied to enhance data diversity and mitigate overfitting.
• Architectures: Two pre-trained architectures, ResNet18 and ResNet50, were implemented. These models were originally trained on the ImageNet dataset.
• Training Strategy: The study utilized transfer learning techniques involving two main strategies:
  1. Feature Extraction: Freezing the initial convolutional layers and training only the newly added fully connected layers.
  2. Fine-tuning: Unfreezing the entire network with a low learning rate to adapt the pre-trained weights to the specific characteristics of brain MRI data.
• Classification: The extracted features were passed through fully connected layers with ReLU activation and dropout regularization. A Softmax classifier was used to output binary probabilities for "Tumor" or "No Tumor."

Key Contributions

• Development of an automated deep learning framework for brain tumor detection using ResNet18 and ResNet50 architectures.
• A comparative analysis of model depth, evaluating the trade-offs between ResNet18 (lighter) and ResNet50 (deeper) in the context of a relatively small medical dataset.
• Application of specific transfer learning strategies, including weight freezing and fine-tuning, to optimize training efficiency and performance.
• Comprehensive evaluation of the impact of preprocessing and data augmentation (resizing, normalization, flipping) on model generalization.

Results
The experimental results demonstrated that the lighter architecture, ResNet18, outperformed the deeper ResNet50 on this specific dataset.

• Accuracy: The optimal configuration (combining resizing, random horizontal flipping, and normalization) yielded a 97% accuracy for ResNet18 and 96% for ResNet50.
• Performance Metrics: ResNet18 achieved a sensitivity of 96.8%, specificity of 97.2%, and an F1-score of 0.97. ResNet50 achieved a sensitivity of 95.9%, specificity of 96.1%, and an F1-score of 0.96.
• Overfitting Analysis: The study observed that ResNet50, with its higher number of parameters, was more prone to overfitting on the limited dataset, resulting in a slightly higher training loss plateau compared to the smoother convergence of ResNet18.
• ROC Analysis: The Receiver Operating Characteristic (ROC) curve indicated that ResNet18 had a slightly larger Area Under the Curve (AUC), confirming its superior ability to differentiate between tumor and non-tumor cases in this context.
• Comparison with Literature: The proposed ResNet18-based approach achieved higher accuracy (97%) compared to recent state-of-the-art methods cited in the literature, such as ResNet50 with Grad-CAM (95.2%) and U-Net with ResNet50 encoder (96.1%).

Significance and Claims
The paper claims that the proposed framework offers a fast, accurate, and cost-effective solution for brain tumor detection. A primary finding is that model depth does not always correlate with better performance in medical imaging tasks with limited data; lighter architectures like ResNet18 can generalize more effectively than deeper ones like ResNet50 by reducing overfitting.

The authors assert that this system supports clinical decision-making by providing a reliable, automated tool that can assist radiologists, reduce manual workload, and potentially improve early detection rates. The study concludes that while the current work focuses on binary classification, the framework lays the groundwork for future extensions into multi-class tumor grading, precise tumor localization, and segmentation using more complex architectures like ResUNet, as well as the integration of explainable AI (XAI) to enhance clinical trust.