Source-linked AI summary
A Deep Learning Approach for Brain Tumor Classification and Segmentation Using a Multiscale Convolutional Neural Network
Francisco Javier Díaz-Pernas, Mario Martínez-Zarzuela, Míriam Antón-Rodríguez, David González-Ortega
TL;DR
The paper targets costly manual brain-tumor localization and classification. It proposes a fully automatic multiscale CNN that jointly segments and classifies three tumor types across MRI views, achieving 0.973 tumor classification accuracy on the evaluated dataset despite documented confusions in non-cerebral regions and sphenoidal sinuses.
Problem
Manual tumor localization and classification is costly and time-consuming, while prior cited studies generally did not combine tumor segmentation with classification by tumor type.
Method
A fully automatic CNN uses three processing pathways at different spatial scales to segment and classify meningioma, glioma, and pituitary tumors in MRI images.
Results
0.973 tumor classification accuracy was obtained, the highest among seven compared approaches using the same dataset.
Takeaways & Limitations
The multiscale CNN provides a joint MRI tumor segmentation and classification approach whose reported performance falls within the top ten methods from the BRATS 2013 benchmark.
Takeaways & Limitations
Non-cerebral areas and sphenoidal-sinus structures produced false positives or tumor-type confusions in example segmentations.
Abstract
from arXiv · showhide
In this paper, we present a fully automatic brain tumor segmentation and classification model using a Deep Convolutional Neural Network that includes a multiscale approach. One of the differences of our proposal with respect to previous works is that input images are processed in three spatial scales along different processing pathways. This mechanism is inspired in the inherent operation of the Human Visual System. The proposed neural model can analyze MRI images containing three types of tumors: meningioma, glioma, and pituitary tumor, over sagittal, coronal, and axial views and does not need preprocessing of input images to remove skull or vertebral column parts in advance. The performance of our method on a publicly available MRI image dataset of 3064 slices from 233 patients is compared with previously classical machine learning and deep learning published methods. In the comparison, our method remarkably obtained a tumor classification accuracy of 0.973, higher than the other approaches using the same database.
1. Introduction
The paper addresses the costly, time-consuming manual localization and classification of brain tumors by developing a deep-learning model for joint segmentation and tumor-type prediction. Its novelty is multiscale CNN processing across three spatial scales, applied to three tumor types and standard MRI views.
- Motivation: Manual tumor localization and classification in large medical-image databases is costly and time-consuming, motivating automatic procedures.The motivation links earlier diagnosis with faster treatment response and improved survival rate.
- Imaging background: MRI is emphasized as the most-used imaging technique because it provides hundreds of high-soft-tissue-contrast 2D slices without ionizing radiation.The paper distinguishes T1, T2, T1-CE, and FLAIR modalities by their tissue contrasts; T1-CE is described as adequate for locating and classifying tumors.
- Background: Brain tumor segmentation commonly differentiates tumor-infected from healthy tissue, often by classifying pixels.Prior work includes both generative methods using anatomical models and discriminative methods learning image features from expert segmentations.
- Related work: Classical machine-learning pipelines typically preprocess images for feature extraction, with skull stripping introducing parameter, prior-information, and computational-time drawbacks.Reported prior accuracies include 79.3%–91.7% for SVM, 75.6%–94.9% for ANN, and 95.3% for a Back Propagation Network.
- Research gap: Earlier cited studies generally segmented tumor regions without also classifying them into different tumor types, leaving joint segmentation and classification as the paper’s objective.The paper contrasts its approach with prior methods using the same three-tumor dataset.
- Contribution: The proposed method uses three CNN processing pathways for three spatial scales, inspired by multiscale Human Visual System processing and intended to extract discriminant texture features.The model targets meningioma, glioma, and pituitary tumors in sagittal, coronal, and axial MRI views.
2. Materials and Methods
The method uses a multiscale, multi-pathway CNN to classify MRI pixels from sliding windows, then aggregates pixel predictions for tumor segmentation and slice-level tumor typing.
- CNN architecture: Each MRI slice is processed pixel by pixel through a multi-pathway CNN that assigns healthy, meningioma, glioma, or pituitary labels.The model uses four output labels: 0 for healthy regions and 1–3 for the three tumor types.
- Multiscale processing: Each pixel is classified from a 65 × 65 neighborhood processed through large-, medium-, and small-scale convolutional pathways.The corresponding kernel sizes are 11 × 11, 7 × 7, and 3 × 3 pixels.
- CNN architecture: The three pathways use two convolutional stages with ReLU and max-pooling, followed by feature concatenation, dropout, and a fully connected classifier.The pathways contain 128, 96, and 64 feature maps at large, medium, and small scales, respectively.
- Dataset: The method operates on 2D slices from 233 patients because clinical CE-MRI acquisitions typically provide sparse slices with large gaps rather than complete 3D volumes.The dataset contains 3064 slices spanning sagittal, coronal, and axial views.
- Training and evaluation: Training uses elastic data augmentation, standardized pixels, 5-fold cross-validation, and sliding-window testing on 612 images per fold.Augmentation doubles the training set to 4904 images per fold, while each fold trains for 80 epochs using SGD.
- Classification and metrics: After pixel labeling, the predicted tumor type is selected from the tumor-label size ratios when the greatest ratio exceeds a confidence threshold; otherwise, the slice can remain nonclassified.Performance is measured with confusion matrices, Dice, sensitivity, and the predicted tumor type ratio score.
3. Results and Discussion
Across 3064 slices evaluated with 5-fold cross validation, the method produced strong segmentation and tumor-classification results, while errors were associated with non-cerebral regions and anatomically confusing structures.
- Segmentation performance: Figure 4 compares predicted and ground-truth tumor regions for meningioma, glioma, and pituitary tumor slices using red, green, and yellow overlays.Red marks predicted regions, green marks ground truth, and yellow marks their intersection.
- Segmentation performance: 0.828 Dice, 0.940 Sensitivity, and 0.967 pttas were the average segmentation metrics across processed slices.Dice was highest for meningiomas and lowest for gliomas, whereas pttas showed the reverse ranking.
- Metric distributions: Sensitivity and pttas values were concentrated near the high end, indicating that almost all processed images achieved high values on these metrics.The Dice distribution was wider, reflecting false positives and false negatives, although Dice remained relatively high.
- Error analysis: Non-cerebral areas and the sphenoidal sinuses produced representative false positives and tumor-type confusions in Figure 6.Examples included non-cerebral tissue labeled as glioma, sinus regions labeled as pituitary tumor, and glioma regions confused with meningioma.
- Tumor classification: 0.973 tumor classification accuracy was obtained from the segmentation output, with the method’s confusion matrix showing very high diagonal values.Classification counts pixel-level tumor-type predictions and selects the highest value above the confidence threshold.
- Confidence threshold: Precision remained above 0.9 at a confidence threshold around 0.92, while thresholds at or below 0.4 were associated with precision close to 100%.The precision curve descended slowly up to 0.88 and then declined sharply; the reported starting accuracy was approximately 0.994.
Comparison with Other Methods
On the same T1-CE MRI dataset, the proposed method outperformed prior feature-driven and deep learning approaches in tumor classification accuracy, while achieving segmentation metrics comparable to leading BRATS methods.
- Segmentation benchmark: BRATS 2013 top-ten methods achieved complete-tumor glioma Dice values between 0.69 and 0.82, providing the benchmark context for segmentation comparison.BRATS uses four MRI modalities and focuses on glioma tumor images.
- Segmentation benchmark: The proposed method’s segmentation performance metrics were reported to be within the range of the BRATS 2013 top-ten methods.The paper compares its segmentation results with a benchmark using glioma images and four MRI modalities.
- Same-dataset classification: 0.973 tumor classification accuracy was achieved by the proposed method, outperforming seven approaches evaluated on the same T1-CE MRI dataset.The comparison included two feature-driven and five deep learning methods.
- Deep learning comparisons: 0.841 accuracy was obtained by Abiwinanda et al., compared with 0.973 for the proposed approach.Abiwinanda et al. selected 700 images per tumor type and did not apply data augmentation.
- Deep learning comparisons: 0.810 and 0.936 accuracies were reported for Pashaei et al.’s CNN and CNN-KELM methods, respectively, both below the proposed method’s 0.973.Their second method substantially improved over their first method.
4. Conclusions
The paper presents a fully automatic multiscale CNN for brain tumor segmentation and classification using publicly available T1-weighted contrast-enhanced MRI data. It reports the highest same-dataset classification accuracy and strong average segmentation metrics despite retaining skull and vertebral-column regions.
- Method: The method is a fully automatic CNN architecture designed for multiscale brain tumor segmentation and classification.Its multiscale design uses three processing pathways.
- Classification results: 0.973 was the highest tumor classification accuracy among the compared approaches using the same dataset.The paper compared its results with seven other classification approaches.
- Segmentation results: 0.828 average Dice, 0.940 average Sensitivity, and 0.967 average pttas were achieved for segmentation.The model successfully segmented and classified meningioma, glioma, and pituitary tumors.
- Scope and caveats: False positives occurred in some images because skull and vertebral-column parts were not removed and the three tumor types were variable.The reported segmentation metrics remained strong despite these conditions.