[Paper Review] Integrating Preprocessing Methods and Convolutional Neural Networks for Effective Tumor Detection in Medical Imaging

Ha Anh Vu|arXiv (Cornell University)|Feb 25, 2024

Brain Tumor Detection and ClassificationNeuroscience3 citations

TL;DR

This study proposes a hybrid approach integrating Gaussian smoothing, bilateral filtering, and K-means clustering for preprocessing brain MRI images, followed by classification using a ResNet50-based CNN. The method achieves 75% accuracy on the Figshare dataset, demonstrating the potential of preprocessing-enhanced deep learning for tumor detection, though limitations in K-means segmentation reduce performance compared to state-of-the-art models.

ABSTRACT

This research presents a machine-learning approach for tumor detection in medical images using convolutional neural networks (CNNs). The study focuses on preprocessing techniques to enhance image features relevant to tumor detection, followed by developing and training a CNN model for accurate classification. Various image processing techniques, including Gaussian smoothing, bilateral filtering, and K-means clustering, are employed to preprocess the input images and highlight tumor regions. The CNN model is trained and evaluated on a dataset of medical images, with augmentation and data generators utilized to enhance model generalization. Experimental results demonstrate the effectiveness of the proposed approach in accurately detecting tumors in medical images, paving the way for improved diagnostic tools in healthcare.

Motivation & Objective

• To improve tumor detection accuracy in medical imaging by enhancing image features through targeted preprocessing techniques.
• To develop and train a deep learning model using a CNN architecture for automated classification of brain tumor types in MRI scans.
• To evaluate the impact of preprocessing—specifically smoothing, edge-preserving filtering, and grayscale conversion—on subsequent model performance.
• To identify limitations in current preprocessing pipelines, particularly K-means clustering, and propose improvements for tumor boundary delineation.
• To benchmark the proposed method against state-of-the-art models on the same dataset for fair performance comparison.

Proposed method

• Preprocessing begins with 7x7 kernel smoothing to reduce noise while preserving structural details in T1-weighted MRI images.
• Bilateral filtering is applied to maintain edge sharpness, especially around tumor boundaries, by considering both spatial proximity and intensity similarity.
• Images are converted to grayscale to simplify analysis and focus on intensity and contrast variations relevant to tissue differentiation.
• K-means clustering with three clusters is used to segment the image into distinct regions, aiming to isolate tumor tissue from normal brain structures.
• A ResNet50-based CNN is trained using data generators and augmentation to improve generalization, with the model optimized for three-class classification: Meningioma, Glioma, and Pituitary tumor.
• The model is trained on 70% of the Figshare dataset (3,064 images) and evaluated on the remaining 30%, with performance tracked via accuracy and loss curves.

Experimental results

Research questions

• RQ1 How do Gaussian smoothing and bilateral filtering affect the detectability of tumor features in preprocessed MRI images?
• RQ2 To what extent does K-means clustering improve tumor segmentation accuracy when used as a preprocessing step for deep learning models?
• RQ3 What is the impact of image preprocessing on the final classification accuracy of a ResNet50-based CNN in brain tumor detection?
• RQ4 How does the performance of the proposed K-means + ResNet50 pipeline compare to state-of-the-art models like InceptionV3 and SVM-based approaches on the same dataset?
• RQ5 What are the key limitations of using K-means clustering in preprocessing for medical image analysis, particularly in tumor boundary detection?

Key findings

• The proposed method achieved a training accuracy of 75%, indicating effective learning on the training set, though test accuracy showed significant variability, suggesting overfitting or generalization issues.
• The model loss decreased overall for both training and test sets, indicating successful error minimization, though a mid-training spike in test loss was observed.
• K-means clustering failed to consistently detect tumors in several test cases, as evidenced by Figures 3, 4, and 5, which show non-detection in Glioma and Pituitary tumor images.
• The method’s 75% accuracy is substantially lower than state-of-the-art models such as Soumik et al. [4] (99%) and Irmak [5] (up to 99.33%), highlighting performance gaps.
• The study identifies K-means clustering as a key limitation in preprocessing, as inaccurate segmentation introduces errors into the training data, reducing model reliability.
• Future work should integrate advanced edge detection techniques to improve tumor boundary delineation and enhance preprocessing fidelity.