Dual-path deep learning framework for accurate and interpretable brain tumor diagnosis

Abstract

Existing frameworks for brain tumor diagnosis often focus on standalone classification or retrieval tasks, limiting clinical interpretability and failing to leverage complementary diagnostic insights. To address this, we propose a novel dual-path deep learning framework that synergistically integrates tumor classification with content-based image retrieval (CBIR). Our approach uniquely combines a lightweight GhostNetV3 backbone with deformable convolutions and a decoupled fully connected (DFC) attention mechanism to simultaneously optimize feature extraction for both tasks. This integration enables dynamic adaptation to irregular tumor morphologies while retrieving visually similar cases, bridging the gap between automated predictions and actionable clinical context. Evaluated on a public T1-weighted contrast-enhanced MRI dataset (233 patients, 3,064 images), the framework achieves state-of-the-art performance: 99.71% classification accuracy (precision/recall/F1 > 0.99) and 97.74% mean average retrieval precision (Prec@10: 99.78%). We further introduce the Classification-Retrieval Agreement Score (CRAS), a novel metric quantifying alignment between classifier predictions and retrieved cases, with a mean score > 0.96 demonstrating robust diagnostic consistency. By enhancing accuracy, interpretability, and computational efficiency, this work advances the clinical viability of AI-driven brain tumor diagnosis.