Source-linked AI summary
Brain Tumor Classification Using Deep Learning Technique -- A Comparison between Cropped, Uncropped, and Segmented Lesion Images with Different Sizes
Ali Mohammad Alqudah, Hiam Alquraan, Isam Abu Qasmieh, Amin Alqudah, Wafaa Al-Sharu
TL;DR
Brain-tumor grading is important after detection for treatment planning. This paper uses a new CNN architecture to classify three tumor types across cropped, uncropped, and segmented MRI images. Uncropped images report the highest overall accuracy and sensitivity among the three image treatments.
Problem
Brain-tumor grading after detection is presented as important for developing an effective treatment plan.
Source: §1 · p. 2Method
A new CNN architecture classifies glioma, meningioma, and pituitary tumors from T1-weighted contrast-enhanced MRI images in uncropped, cropped, and segmented datasets.
Source: §3.1 · p. 3, §6 · p. 7Results
Uncropped images had 99% Accuracy and 98.85% Sensitivity, versus cropped images at 98.40% Accuracy and 98.18% Sensitivity and segmented images at 97.62% Accuracy and 97.40% Sensitivity.
Source: §5 · p. 6Takeaways & Limitations
The authors report high grading performance for all three tumor classes across uncropped, cropped, and segmented image cases.
Source: §6 · p. 7The conclusion suggests that adding more MR images with different weights and contrast-enhancement techniques may further improve generalizability and robustness.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Deep Learning is the newest and the current trend of the machine learning field that paid a lot of the researchers' attention in the recent few years. As a proven powerful machine learning tool, deep learning was widely used in several applications for solving various complex problems that require extremely high accuracy and sensitivity, particularly in the medical field. In general, brain tumor is one of the most common and aggressive malignant tumor diseases which is leading to a very short expected life if it is diagnosed at higher grade. Based on that, brain tumor grading is a very critical step after detecting the tumor in order to achieve an effective treating plan. In this paper, we used Convolutional Neural Network (CNN) which is one of the most widely used deep learning architectures for classifying a dataset of 3064 T1 weighted contrast-enhanced brain MR images for grading (classifying) the brain tumors into three classes (Glioma, Meningioma, and Pituitary Tumor). The proposed CNN classifier is a powerful tool and its overall performance with accuracy of 98.93% and sensitivity of 98.18% for the cropped lesions, while the results for the uncropped lesions are 99% accuracy and 98.52% sensitivity and the results for segmented lesion images are 97.62% for accuracy and 97.40% sensitivity.
1. INTRODUCTION
Brain tumors vary by type and malignancy, and grading is presented as important after detection for treatment planning. The study evaluates a new CNN architecture on cropped, uncropped, and segmented brain-tumor images.
1. INTRODUCTION
- Glioma, meningioma, and pituitary tumors are distinguished by their locations in glial tissue and spinal cord, protective membranes, and pituitary gland, respectively.The paper describes brain tumors as abnormal brain-cell growth affecting the central nervous system.Source: §1 · p. 1
- MRI and CT support initial tumor evaluation, while suspected cases requiring more diagnostic information may need surgical biopsy.The introduction also describes computer-aided diagnosis systems as a way to help physicians improve early cancer detection accuracy.Source: §1 · p. 1
- The study applies a new CNN architecture to grade three tumor classes using cropped, uncropped, and segmented T1-weighted contrast-enhanced MRI images.The authors identify a new architecture rather than transfer learning or a pretrained CNN as a contribution.Source: §1 · p. 2
- Brain-tumor grading is described as a critical step after detection for developing an effective treatment plan.The introduction notes that higher-grade diagnosis is associated with a very short expected life.Source: §1 · pp. 1–2
2. LITERATURE REVIEW
The literature review describes brain-tumor classification systems combining image processing, feature extraction, and machine-learning classifiers. Reported approaches span different tumor classes, datasets, and performance levels.
2. LITERATURE REVIEW
- Prior systems combine segmentation or ROI processing with feature extraction and classifiers, including DWT with PCA and DNN, and CNN with GLCM features.Other reviewed approaches include fuzzy C-means segmentation followed by texture and shape features, SVM, or DNN classification.Source: §2 · p. 2
- Reported accuracy included 82.27% for a two-feature CNN system on the Gl-Pt dataset and 98% using four selected wavelet features with an ANN.The review also reports 96.97% accuracy and 97.0% sensitivity for a DWT/deep-learning system.Source: §2 · p. 2
- ROI augmentation and fine ring-form partition increased reported accuracy for intensity-histogram, GLCM, and BoW methods.The review gives respective changes of 71.39% to 78.18%, 83.54% to 87.54%, and 89.72% to 91.28%.Source: §2 · p. 2
3. MATERIAL AND METHODS
The study aims to develop a CNN architecture for grading brain tumors from T1-weighted contrast-enhanced MRI. It uses a dataset with three tumor classes and lesion masks that support cropped-region inputs.
3. MATERIAL AND METHODS
- The study aims to provide a new CNN architecture for grading brain tumors using T1-weighted contrast-enhanced MR images.Figure 1 is identified as a block diagram of the proposed methodology.Source: §3 · pp. 2–3, Figure 1 · p. 3
3.1 Dataset
- The dataset contains 3,064 T1-weighted contrast-enhanced MRI images labeled as glioma, meningioma, or pituitary tumor.The dataset includes patient identifiers, tumor masks, tumor borders, and class labels; the lesion mask is used to crop the tumor ROI.Source: §3.1 · p. 3, Table 1 · p. 3
3.2 Methodology
The methodology adapts an 18-layer CNN for three image treatments and evaluates its grading predictions against image labels using confusion-matrix metrics. Training accuracy is also tracked across image cases and input sizes.
3.2.1 Convolutional Neural Network (CNN)
- CNNs use convolution with trainable filters and downsampling, and the described network includes input, convolution, ReLU, fully connected, classification, and output layers.Figure 2 illustrates uncropped, cropped, and segmented image examples.Source: §3.2.1 · p. 3, Figure 2 · p. 3
- The dataset is partitioned for training and testing, with the methodology passage specifying 70%, 15%, and 15% for the three splits.Figure 4 tracks accuracy variation during training and validation across three image cases and different input sizes.Source: §3.2.1 · p. 3, §3.2.3 · p. 4, Figure 4 · p. 4
3.2.2 Proposed CNN Architecture
- The proposed CNN is an 18-layer architecture modified from an architecture previously used for OCT image classification.It is applied to cropped, uncropped, and segmented image datasets.Source: §3.2.2 · p. 3
3.2.3 Performance Evaluation
- The evaluation compares CNN outputs with original image labels using confusion matrices to calculate accuracy, sensitivity, precision, and specificity.The stated confusion-matrix indices are true positive, false positive, false negative, and true negative.Source: §3.2.3 · pp. 3–4
4.1 Results of Uncropped Lesion Images
For uncropped lesion images, accuracy was 97.4% at 128x128, 99.0% at 64x64, and 99.2% at 32x32.
4.1 Results of Uncropped Lesion Images
- 99.2% accuracy was reported for uncropped lesion images sized 32x32.Source: §4.1 · p. 4
- 99.0% accuracy was reported for uncropped lesion images sized 64x64.Source: §4.1 · p. 4
- 97.4% accuracy was reported for uncropped lesion images sized 128x128.Source: §4.1 · p. 4
4.2 Results of Cropped Lesion Images
For cropped lesion images, accuracy was 97.4% at 128x128, 98.4% at 64x64, and 96.9% at 32x32.
4.2 Results of Cropped Lesion Images
- 98.4% accuracy was reported for cropped lesion images sized 64x64.Source: §4.2 · p. 4
- 97.4% accuracy was reported for cropped lesion images sized 128x128.Source: §4.2 · p. 4
- 96.9% accuracy was reported for cropped lesion images sized 32x32.Source: §4.2 · p. 4
4.3 Results of Segmented Lesion Images
For segmented lesion images, accuracy was 97.5% at 128x128, 97.6% at 64x64, and 98.4% at 32x32.
4.3 Results of Segmented Lesion Images
- 98.4% accuracy was reported for segmented lesion images sized 32x32.Source: §4.3 · p. 4
- 97.6% accuracy was reported for segmented lesion images sized 64x64.Source: §4.3 · p. 4
- 97.5% accuracy was reported for segmented lesion images sized 128x128.Source: §4.3 · p. 4
4. RESULTS
Experiments used cropped and uncropped datasets with 64-image minibatches, ADAM, and a 10-3 initial learning rate, yielding 1600 iterations on a desktop computer.
5. DISCUSSION
Across three tumor classes, uncropped images had the strongest overall performance, and the paper reports that its method outperformed prior multiclass studies. Performance was assessed across three input sizes and three image treatments.
5. DISCUSSION
- Uncropped images achieved 99% accuracy, 99.64% specificity, 98.85% sensitivity, and 98.98% precision overall.The performance evaluation figures cover input sizes 128, 64, and 32.Source: §5 · p. 6, Figure 8 · p. 6, Figure 9 · p. 6
- Cropped images had 98.40 accuracy and 98.18% sensitivity, while segmented images had 97.62% accuracy and 97.40% sensitivity.The paper also reports 99.19% specificity and 98.19% precision for cropped images, and 98.78% specificity and 97.43% precision for segmented images.Source: §5 · p. 6, Figure 8 · p. 6, Figure 9 · p. 6
- The authors attribute weaker cropped-image results to discarded surrounding pixels and weaker segmented-image results to black mask pixels and loss of lesion color texture.They state that lesion texture color could describe the lesion and that black mask pixels are counted even though they are not present in the lesion.Source: §5 · p. 6
- The method addresses three-class classification of glioma, meningioma, and pituitary tumors, and the authors report outperforming prior related studies.They state that most listed methods exceeded 90% recognition, with the highest at 98%, and claim their method outperforms them by a margin.Source: §5 · pp. 4, 7
6. CONCLUSION
The paper presents a CNN architecture for classifying brain tumors across uncropped, cropped, and segmented image datasets. It reports high accuracy and sensitivity across these cases and suggests that broader imaging data could support potential generalization and robustness.
6. CONCLUSION
- The proposed CNN classifies glioma, meningioma, and pituitary tumors across uncropped, cropped, and segmented ROI datasets, with high accuracy and sensitivity in all cases.The input is T1-weighted contrast-enhanced brain MR images.Source: §6 · p. 7
- The authors suggest that adding more brain MR images with different weights and contrast-enhancement techniques may support potential generalization and robustness on larger image databases.Source: §6 · p. 7
- Table 2 compares accuracy results for related-work methods and the proposed method.Source: Table 2 · p. 7