[Paper Review] Deep Learning in Medical Image Classification from MRI-based Brain Tumor Images

Xiaoyi Liu, Zhuoyue Wang|arXiv (Cornell University)|Aug 1, 2024

Brain Tumor Detection and Classification5 citations

TL;DR

The paper evaluates four pre-trained CNNs on brain tumor MRI classification and introduces a MobileNet-BT model that unfreezes all layers for superior accuracy (0.9924).

ABSTRACT

Brain tumors are among the deadliest diseases in the world. Magnetic Resonance Imaging (MRI) is one of the most effective ways to detect brain tumors. Accurate detection of brain tumors based on MRI scans is critical, as it can potentially save many lives and facilitate better decision-making at the early stages of the disease. Within our paper, four different types of MRI-based images have been collected from the database: glioma tumor, no tumor, pituitary tumor, and meningioma tumor. Our study focuses on making predictions for brain tumor classification. Five models, including four pre-trained models (MobileNet, EfficientNet-B0, ResNet-18, and VGG16) and one new model, MobileNet-BT, have been proposed for this study.

Motivation & Objective

• Classify four brain tumor/mri conditions: glioma, pituitary tumor, meningioma, and no tumor.
• Compare pre-trained CNNs (MobileNetV2, ResNet-18, EfficientNet-B0, VGG16) on brain tumor MRI data.
• Propose and evaluate a customized MobileNet-based model (MobileNet-BT) with full fine-tuning.
• Analyze the impact of data augmentation and training strategies on model performance.

Proposed method

• Use Brain Tumor MRI Dataset with 7023 images distributed as 2000 no tumor, 1757 pituitary, 1621 glioma, 1645 meningioma.
• Apply image augmentation (horizontal/vertical flips, ±10° rotation) and resize to 224x224.
• Train with transfer learning; freeze all but last layer for four baseline models; evaluate with accuracy, F1-score, precision, recall, and average loss.
• Unfreeze all MobileNetV2 layers and replace final classifier to create MobileNet-BT with dropout and two dense layers for four-class output.
• Use cross-entropy loss and a learning rate starting at 0.001 with a scheduler multiplying by 0.1 every 8 epochs; stop when validation accuracy stalls for 8 epochs.

Experimental results

Research questions

• RQ1 How do standard pre-trained CNNs (MobileNetV2, ResNet-18, EfficientNet-B0, VGG16) perform on MRI-based brain tumor classification?
• RQ2 Does fully unfreezing MobilenetV2 and customizing a classifier (MobileNet-BT) improve classification accuracy and F1-score for brain tumor MRI images?
• RQ3 What is the impact of data augmentation and learning-rate scheduling on model performance in this medical imaging task?
• RQ4 How do the models compare in terms of accuracy, precision, recall, F1-score, and loss on the brain tumor dataset?

Key findings

• MobileNetV2 achieved 0.8445 accuracy and 0.8431 F1-score with 3.5M parameters.
• ResNet-18 achieved 0.8659 accuracy and 0.8635 F1-score with 11.7M parameters.
• EfficientNet-B0 achieved 0.8933 accuracy and 0.8919 F1-score with 5.3M parameters.
• VGG16 achieved 0.9497 accuracy and 0.9494 F1-score with 138M parameters.
• MobileNet-BT achieved 0.9924 accuracy and 0.9924 F1-score, outperforming all baselines; best results at epoch 10 vs around epoch 20 for others.