Few-shot brain tumor classification: meta- vs metric-learning comparison

Abstract

Medical imaging requires accurate brain tumor recognition because precise classification is essential for early diagnosis and effective treatment planning. A major challenge in medical applications is that deep learning models typically require extensive amounts of labeled data to perform well. To address this, this research evaluates three few-shot learning (FSL) approaches-prototypical networks, Siamese networks, and model-agnostic meta-learning (MAML)-for brain tumor classification using the Figshare brain tumor dataset. The results show that prototypical networks consistently outperform the other approaches, achieving 89.07% accuracy (95% CI: 88.12–89.96%), 88.73% precision, and 88.67% recall, making them the optimal solution for this task. Siamese networks achieve 83.73% accuracy (95% CI: 82.64–84.76%), while MAML demonstrates significantly reduced performance, with 43.70% accuracy (95% CI: 42.10–45.22%). This study demonstrates that FSL can be applied effectively for medical image classification, with prototypical networks achieving the best performance in brain tumor detection. The inclusion of confidence intervals further validates the robustness and reliability of the results. Future research will focus on improving feature representation and exploring hybrid approaches to better handle rare tumor classes, thereby enhancing the clinical applicability of FSL models.