Accurate and secure classification of brain tumors from magnetic resonance imaging remains a critical problem in medical image analysis. Deep convolutional models deliver strong predictive performance but often require large, diverse datasets and offer limited interpretability. At the same time, radiologists face significant diagnostic challenges due to subtle textural variations, heterogeneous tumor morphology, and restricted access to multiparametric imaging. These clinical difficulties are further compounded by strict privacy regulations that limit the sharing of patient images. To address these issues, we propose a hybrid framework that combines mathematically defined Whirling Triangle partitioning for localized handcrafted feature extraction, convolutional neural network-based image embeddings, and a Cheon–Kim–Kim–Song homomorphic encryption–enabled inference pipeline demonstrating secure predictions without revealing plaintext data. The Whirling Triangle decomposition produces multi-scale triangular subregions from which we extract intensity, gradient, and textural descriptors (including Local Binary Patterns and Gray-Level Co-occurrence Matrix), which are fused with deep image features to produce a compact, discriminative representation. Experiments were conducted on a dataset of 7,023 T1-weighted contrast-enhanced axial brain tumor Magnetic Resonance Imaging (MRI) images, which shows 98.0% accuracy in the plaintext setting and 96.7% accuracy under encrypted inference.
Dutta et al. (Thu,) studied this question.