GeoMedNet

# CESEN-GPARS: Interpretable Deep Learning for Medical Image Classification ## 🧠 Overview **CESEN-GPARS** is an interpretable deep learning framework for medical image classification. It is designed to extract non-linear features and preserve structural consistency across complex high-dimensional data. This framework introduces two key components: - **CESEN**: Curvature-Enhanced Signal Encoding Network- **GPARS**: Geometric Profile-Aligned Regularization Strategy Together, they provide curvature-aware representation and geometry-aligned optimization for robust and explainable classification in medical imaging. ## 🔍 Key Features - 💡 **Curvature-guided Feature Extraction**: Enhances local geometry using directional curvatures and angular modulation.- 🧩 **Geometric Profile Alignment**: Encourages global structural consistency via optimization-based surface fitting.- 🧠 **Interpretability-Preserving**: Maintains semantic structure and enhances visualization potential for medical diagnosis.- 📈 **Superior Generalization**: Outperforms state-of-the-art on ImageNet, Caltech-256, Oxford Flowers, and DTD datasets. ## 🧱 Architecture The framework consists of:- A backbone encoder with curvature-enhanced modules (CESEN)- A geometry-aware loss function (GPARS)- Classification head with semantic alignment 📌 *For details, refer to Figures and diagrams on pages 5–8 of the paper.* ## 📦 Project Structure# CESEN-GPARS: 可解释深度学习用于医学图像分类 ## 🧠 概述 **CESEN-GPARS**是一款面向医学图像分类任务的可解释深度学习框架，旨在提取非线性特征，并在复杂高维数据中保留结构一致性。该框架引入两大核心组件： - **CESEN**：曲率增强信号编码网络（Curvature-Enhanced Signal Encoding Network） - **GPARS**：几何轮廓对齐正则化策略（Geometric Profile-Aligned Regularization Strategy） 二者协同为医学成像领域的鲁棒且可解释的分类任务提供曲率感知表征与几何对齐优化手段。 ## 🔍 核心特性 - 💡 **曲率引导特征提取**：通过方向曲率与角度调制增强局部几何信息。 - 🧩 **几何轮廓对齐**：借助基于优化的曲面拟合方法，推动全局结构一致性。 - 🧠 **可解释性保障**：保留语义结构，提升医学诊断所需的可视化潜力。 - 📈 **泛化性能优异**：在ImageNet、Caltech-256、牛津花卉数据集（Oxford Flowers）与可描述纹理数据集（DTD）上的表现优于当前前沿方法。 ## 🧱 架构 该框架包含： - 搭载曲率增强模块（CESEN）的骨干编码器 - 几何感知损失函数（GPARS） - 带有语义对齐功能的分类头 📌 *详细内容请参阅论文第5至8页的图表与示意图。* ## 📦 项目结构