Source-linked AI summary
Brain tumor MRI image classification with feature selection and extraction using linear discriminant analysis
V. P. Gladis Pushpa Rathi, S. Palani
TL;DR
Brain MRI tumor diagnosis is challenging because tumors vary in size, shape, location, and intensity. The paper combines intensity, shape, and texture features with LDA-based selection; it reports 98.87% classification accuracy using all features with LDA.
Problem
Early, accurate brain-tumor diagnosis is important for treatment planning, but image diagnosis is challenging because tumors vary in size, shape, location, and intensity.
Source: §1 · p. 2Method
The method extracts intensity, shape, and texture features, selects vital features using LDA, and compares results with PCA-based dimension reduction.
Source: §1 · p. 2, §5 · p. 14Results
98.87% classification accuracy was reported using all features with LDA.
Source: §4 · p. 10Takeaways & Limitations
The authors conclude that selected intensity, shape, and texture features support brain-tumor MRI classification, with vital features selected using LDA.
Source: §5 · p. 14The described approach can ignore correlations among features and select redundant features.
Source: yes · p. 6
Abstract
from arXiv · showhide
Feature extraction is a method of capturing visual content of an image. The feature extraction is the process to represent raw image in its reduced form to facilitate decision making such as pattern classification. We have tried to address the problem of classification MRI brain images by creating a robust and more accurate classifier which can act as an expert assistant to medical practitioners. The objective of this paper is to present a novel method of feature selection and extraction. This approach combines the Intensity, Texture, shape based features and classifies the tumor as white matter, Gray matter, CSF, abnormal and normal area. The experiment is performed on 140 tumor contained brain MR images from the Internet Brain Segmentation Repository. The proposed technique has been carried out over a larger database as compare to any previous work and is more robust and effective. PCA and Linear Discriminant Analysis (LDA) were applied on the training sets. The Support Vector Machine (SVM) classifier served as a comparison of nonlinear techniques Vs linear ones. PCA and LDA methods are used to reduce the number of features used. The feature selection using the proposed technique is more beneficial as it analyses the data according to grouping class variable and gives reduced feature set with high classification accuracy.
1. INTRODUCTION
The introduction frames brain MRI classification as difficult because tumor characteristics vary and image gray values can shift with enhancement or noise. It motivates reducing complex image data to selected features, then describes a PCA–LDA reduction approach.
1. INTRODUCTION
- Early, accurate brain-tumor diagnosis matters for treatment planning, but image variation in tumor size, shape, location, and intensity makes diagnosis challenging.The passage says diagnosis can only be performed by professional neuroradiologists.Source: §1 · p. 2
- MRI supports non-invasive soft-tissue imaging, but gray values can change with over-enhancement or noise.The paper identifies statistical, intensity, symmetry, and texture features as inputs used to classify tumors.Source: §1 · p. 2
- Feature extraction reduces raw image data to a smaller feature representation, while feature selection removes irrelevant or redundant features.The introduction associates these reductions with lower memory and computation demands and reduced risk of overfitting.Source: §1 · p. 2
- PCA retains all eigenvalues initially; the two largest account for 99.99% of their total, followed by LDA reduction from 15 features to 2.The authors state that this combined process increases accuracy and decreases complexity and computational time.Source: §1 · pp. 2–3
2. RELATED WORKS
The related-work section surveys computer-assisted tumor analysis and several segmentation or grading approaches. It then presents PCA–LDA feature reduction with SVM classification as the paper’s approach to MRI classification.
2. RELATED WORKS
- Computer-assisted methods are motivated by a need to retain diagnostic accuracy while improving interoperability and transparency in clinical decision-making.The passage notes that different histopathologists may reach different outputs from the same recognized features.Source: §2 · p. 3
- Prior work includes fuzzy cognitive maps for brain-tumor grading, reporting 90.26% accuracy for low-grade and 93.22% for high-grade tumors.That approach is described as addressing tumor characterization and grade determination.Source: §2 · p. 3
- Other cited approaches cover MRI abnormality segmentation with mosaiced images, tumor-growth and volume measurement, and automatic segmentation or tissue segmentation.The passage attributes these approaches to Shafab Ibrahim and Noor Elaiza, S. Karpagam and S. Gowri, Matthew C. clrk Lawrence et al., and Carlos A. Patta et al..Source: §2 · pp. 3–4
- The paper combines PCA and LDA for feature reduction and uses SVM to classify MRI images.The authors characterize their reported accuracy as high relative to the previous work discussed.Source: §2 · p. 4
3. PROPOSED METHOD
The proposed method processes brain MRI data through normalization, feature extraction and selection, and classification. It extracts intensity, shape, and texture features, then applies feature-selection procedures to reduce the set.
3. PROPOSED METHOD
- The system architecture comprises a brain-tumor database, normalization, feature selection, feature extraction, and classification.Figure 1 is identified as the architecture of the proposed method.Source: §3 · p. 4, Figure 1 · p. 4
3.1. Data Description
- The experiments use MRI volumes from 20 glioma patients, each with T1, T2, and FLAIR sequences and 24 axial slices at 5 mm thickness.Images were acquired on 3.0T Siemens devices; each feature set was normalized to 0–255.Source: §3.1 · p. 5, Table 1 · p. 5
3.2. Normalization
- MRI images are normalized to gray-level values from 0 to 1 before feature extraction.The authors state that normalization reduces the intensity dynamic range and simplifies feature extraction.Source: §3.2 · p. 5
3.3. Feature Extraction
- Feature extraction produces intensity, shape, and texture descriptors, which may contain redundancy for later selection.The listed descriptors include intensity statistics, shape measures, and texture measures.Source: §3.3 · p. 5
3.4. Feature Selection
- Feature selection keeps a subset of available features, aiming for the fewest dimensions that contribute to high accuracy.The remaining dimensions are discarded as unimportant.Source: §3.4 · p. 6
3.4.1. Forward Selection
- Forward selection starts with no variables and adds features one by one according to a two-tailed t-test of differences between class means.The criterion evaluates each feature’s discriminative power for separating two classes.Source: §3.4.1 · p. 6
yes
This section describes feature selection and classification methods, including PCA and LDA dimensionality reduction and SVM classification.
yes
- Two-tailed t-test ranks discriminative power, then p<0.1 filtering precedes subset selection; Figure 2 depicts forward-selection steps.Source: yes · pp. 6–7, Figure 2 · p. 6
- The t-test assumes normal features with unknown but equal variances; ignoring correlations can retain redundant features and affect classification.Source: yes · pp. 6–7
3.4.2. Backward Selection
- Backward selection removes features one at a time using SVM leave-one-out error-bound changes, with validation error measured separately from training error.Source: §3.4.2 · p. 7
3.5. Classification
- Classification trains on known data and tests on unknown data, with accuracy depending on training efficiency.Source: §3.5 · p. 7
3.5.1. Principal Component Analysis
- PCA projects inputs onto eigenvectors associated with the largest eigenvalues, reducing an n-dimensional input to an m-dimensional feature vector where m is much smaller than n.Source: §3.5.1 · p. 8
3.5.2. Linear Discriminant Analysis
- LDA forms linear combinations that maximize differences between class means, using within-class and between-class scatter measures.Source: §3.5.2 · pp. 8–9
2.5.2. Support Vector Machine
- SVM maps inputs to higher-dimensional features for a maximum-margin hyperplane; KKT conditions solve constrained optimization, and nonzero solutions identify support vectors.Source: §2.5.2 · p. 9
4. EXPERIMENT RESULTS
The experiments report feature distributions and classification results, and compare SVM classification with and without continuous training.
4. EXPERIMENT RESULTS
- Among 60 selected features, 22 are intensity-based, 5 shape-based, and 33 texture-based.The authors report all three feature types as useful for classification.Source: §4 · p. 10, Table 2 · p. 10
- The selected-feature distribution across sequences is T1: 10, T2: 20, and FLAIR: 30; the authors report FLAIR provides the most information for tumor segmentation.They describe this ordering as consistent with greater pathological-discrimination sensitivity for FLAIR and T2 than T1.Source: §4 · p. 10
- 97.82% average correct rate, with 1.0% FP and 2.50% FN; all features with LDA achieve 98.87%, while four PCA components classified with LDA and SVM achieve 96%.Figure 4 presents overall accuracy performance.Source: §4 · pp. 10, 12
- The preprocessing results show original, blurred, edge-detected, segmented, and normalized images.The figure caption also describes the final image as showing all processing steps.Source: §4 · p. 11
- Continuous SVM training gives more identification of similar properties than the reported alternative without continuous training.The experiment presents outputs for both training conditions.Source: §4 · pp. 12–13
5. CONCLUSIONS
The conclusion describes a brain-MRI classification pipeline combining feature extraction, LDA-based selection, and SVM classification, and reports 98.87% accuracy.
5. CONCLUSIONS
- The proposed pipeline collects and normalizes images, extracts intensity, shape, and texture features, selects vital features using LDA, and classifies them.The conclusion compares the results with PCA dimensionality reduction.Source: §5 · p. 14
- The conclusion reports 98.87% classification accuracy by SVM alongside features selected or extracted by PCA.Source: §5 · p. 14
- The authors report training with and without continuous data to minimize error and increase accuracy, and expect adding fMRI and image moments to improve results.Source: §5 · p. 14