NeuroLesionNet

# NeuroLesionLoc: Object Detection & Lesion Localization in Neuroimaging > A hybrid AI framework for interpretable and robust lesion localization across multimodal brain imaging datasets. ## 🧠 Overview **NeuroLesionLoc** is a state-of-the-art computational pipeline for the detection and localization of pathological lesions in neuroimaging. It integrates anatomical priors, structured latent representations, and domain-adaptive inference mechanisms to improve both performance and interpretability in clinical neuroimaging tasks such as stroke, tumor, and trauma detection. This repository accompanies the research paper: **"Object Detection Lesion Localization in Neuroimaging: Enhancing Diagnostic Accuracy in Brain Disorders"** *Lei Zhao, Zhejiang University of Technology* ## 📌 Key Features - **NeuroManifold Encoder (NME)**: A variational encoder that embeds multimodal brain imaging data into a biologically consistent, low-dimensional manifold. - **Cortex-guided Adaptive Projection (CAP)**: An inference mechanism that uses anatomical priors and uncertainty-aware reweighting to refine lesion localization. - **Compositional Spatial Decoder**: Reconstructs input data through interpretable basis functions and attention-driven modulation. - **Auxiliary-Aware Modulation**: Integrates subject-specific attributes (e.g., age, sex, diagnosis) to enhance latent space alignment and diagnostic relevance. - **Cross-domain Alignment**: Employs Maximum Mean Discrepancy (MMD) to harmonize feature distributions across datasets and institutions. ## 🧪 Datasets Used - **ISLES 2022**: Stroke lesion segmentation with multimodal MRI.- **Brain Tumor Dataset**: T1, T2, FLAIR MRI scans for tumor detection (glioma, meningioma, pituitary).- **BraTS 2021**: Glioma segmentation with multimodal MRI and survival prediction.- **ATLAS**: Stroke lesion mapping with T1-weighted MRIs and rich metadata. ## 🧬 Model Architecture The pipeline comprises:1. Graph-Regularized Latent Encoding using anatomical adjacency graphs.2. Spatial-Temporal Compositional Decoder with attention-modulated basis functions.3. Anatomical projection and uncertainty-aware correction.4. Optional classification head for diagnostic tasks. > See Figures 1–4 in the paper for detailed diagrams of the architecture. ## 📈 Performance Highlights | Dataset | Accuracy | F1 Score | AUC ||----------------|----------|----------|---------|| ISLES 2022 | 93.61% | 91.72% | 94.34% || Brain Tumor | 92.25% | 90.83% | 93.15% || BraTS 2021 | 94.21% | 92.03% | 95.42% || ATLAS | 93.07% | 90.85% | 94.13% | > Outperforms YOLOv5, RetinaNet, DETR, CenterNet and others across all datasets. ## 📦 Installation ```bashgit clone https://github.com/yourusername/NeuroLesionLoc.gitcd NeuroLesionLocpip install -r requirements.txt