Trade-Off Analysis of Classical Machine Learning and Deep Learning Models for Robust Brain Tumor Detection: Benchmark...

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computational neuroscience

Background:

• Brain tumor detection relies heavily on medical image analysis.
• Training deep learning models typically requires extensive labeled datasets, which are costly and time-consuming to acquire.
• This study investigates the necessity of deep learning for small medical datasets and the potential of self-supervised learning to mitigate annotation costs.

Purpose of the Study:

• To comparatively analyze traditional machine learning and deep learning models, including self-supervised approaches, for brain tumor classification using small medical image datasets.
• To evaluate the robustness, transferability, and generalization capabilities of these models on unseen data across different domains.
• To understand the trade-offs between model complexity, annotation requirements, and practical deployment in medical imaging.

Main Methods:

• Four distinct models were evaluated: Support Vector Machine with Histogram of Oriented Gradients (SVM+HOG), ResNet18 (CNN), Vision Transformer (ViT-B/16), and Simple Contrastive Learning of Visual Representations (SimCLR) for self-supervised learning.
• A dataset of 2870 brain MRI images across four classes (glioma, meningioma, pituitary, nontumor) was utilized.
• Models were trained with consistent data augmentation and early stopping, with performance assessed using accuracy, precision, recall, F1-score, and convergence on both within-domain and cross-domain unseen test data.

Main Results:

• ResNet18 demonstrated superior performance with the highest validation accuracy (99.77%) and strong cross-domain accuracy (95%).
• Vision Transformer (ViT-B/16) achieved 98% within-domain and 93% cross-domain test accuracy.
• Self-supervised SimCLR reached 97% within-domain and 91% cross-domain accuracy, showcasing its potential for reduced annotation.
• SVM+HOG provided competitive within-domain accuracy (97%) but significantly lower cross-domain performance (80%).

Conclusions:

• Meaningful trade-offs exist between model complexity, annotation effort, and deployment feasibility in medical imaging.
• Deep learning models, particularly CNNs like ResNet18, offer high performance for brain tumor classification.
• Self-supervised learning presents a viable alternative for reducing annotation costs while maintaining competitive performance.