XMorph: Explainable Brain Tumor Analysis Via LLM-Assisted Hybrid Deep Intelligence

The official implementation of XMorph, a clinically interpretable and computationally efficient framework for fine-grained brain tumor classification. XMorph bridges the gap between high-performance deep learning and clinical trust by fusing deep visual features, nonlinear dynamics (IWBN), and quantitative radiological biomarkers with dual-channel explainability.

---

📊 Performance Summary

Validated results on the Figshare and BraTS datasets as detailed in the accompanying paper:

Metric	Result
Classification Accuracy	96.0%
Segmentation Dice Score (WT)	0.932
Interpretability	Dual-Channel (Visual + Textual)

---

🔍 Capability Comparison

How XMorph differs from existing state-of-the-art diagnostic tools:

Feature / Capability	Deepak & Ameer [4]	Cheng [21]	Mahesh et al. [7]	Saeed et al. [8]	Sultan et al. [13]	Temtam et al. [14]	Rashed et al. [9]	XMorph (Ours)
Deep Feature Learning	✅	✅	✅	✅	✅	✅	✅	✅
Fractal Dimension (FD)	❌	❌	❌	❌	✅	✅	❌	✅
Chaotic Metrics (ApEn, LE)	❌	❌	❌	❌	❌	❌	❌	✅
IWBN (Boundary Enhancement)	❌	❌	❌	❌	❌	❌	❌	✅
Clinical Biomarkers (REI, MLS, Dskull)	❌	❌	❌	❌	❌	❌	❌	✅
Visual XAI (Heatmaps)	❌	❌	✅	✅	✅	❌	❌	✅
Textual XAI (LLM Rationales)	❌	❌	❌	❌	❌	❌	✅	✅

References:
[4] S. Deepak and P. Ameer, "Brain tumor classification using deep cnn features via transfer learning," Computers in Biology and Medicine, vol. 111, p. 103345, 2019.
[7] T. R. Mahesh et al., "An xai-enhanced efficientnetb0 framework for precision brain tumor detection in mri imaging," Journal of Neuroscience Methods, vol. 410, p. 110227, 2024.
[8] T. Saeed et al., "Neuro-xai: Explainable deep learning framework based on deeplabv3+ and bayesian optimization for segmentation and classification of brain tumor in mri scans," Journal of Neuroscience Methods, vol. 410, p. 110247, 2024.
[9] E. A. Rashed et al., "Automatic generation of brain tumor diagnostic reports from multimodality mri using large language models," 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), 2025.
[13] H. Sultan et al., "Estimation of fractal dimension and segmentation of brain tumor with parallel features aggregation network," Fractal and Fractional, vol. 8, no. 6, p. 357, 2024.
[14] A. Temtam, L. Pei, and K. Iftekharuddin, "Computational modeling of deep multiresolution-fractal texture and its application to abnormal brain tissue segmentation," arXiv preprint arXiv:2306.04754, 2023.
[21] J. Cheng, "Brain tumor dataset," 2017. [Online]. Available: https://doi.org/10.6084/m9.figshare.1512427.v8

---

⚙️ Pipeline Stages

1. Stage 1 – Automated Tumor Segmentation

   • Input: Raw CE-T1 MRI Image.
   • Process: DeepLabV3-based semantic segmentation using a ResNet-50 backbone.
   • Output: Binary tumor mask and boundary contour.

2. Stage 2 – Tumor-Specific & IWBN Features

   • Input: Tumor Mask + Boundary Contour.
   • Process: Extraction of radiological clinical features (REI, MLS) and our novel Information-Weighted Boundary Normalization (IWBN) time-series and Non-linear features.
   • Output: Quantitative feature arrays (Non_Linear_Features.npy, information_weighted_time_series.npy,clinical_features.npy).

3. Stage 3–5 – Feature Fusion and Classification

   • Input: Deep CNN Embeddings + Stage 2 Feature Vectors.
   • Process: PCA-based dimensionality reduction followed by synergistic fusion and classification via an optimized XGBoost model.
   • Output: Predicted tumor class (Glioma, Meningioma, Pituitary) and confidence scores.

4. Stage 6 – Dual-Channel Explainability

   • Input: Model Weights + SHAP values of fused features.
   • Process: Generation of visual Grad-CAM++ saliency maps and LLM-assisted diagnostic narratives (GPT-5).
   • Output: Interpretable visual heatmaps and textual clinical rationales.

---

📝 Notes

Reproducibility: All experiments use fixed random seeds (See Scripts).

LLM Stage: Textual explanations are exported as CSV files and can be re-processed with GPT-4 or GPT-5 for deterministic narrative reproduction. 🚀 Setup & Reproducibility Notebooks must be run sequentially to maintain the data dependency chain. Follow these steps to set up your environment: code Bash

Create and activate virtual environment

python -m venv .venv source .venv/bin/activate

Install dependencies

jupyter notebook

• Execution Order: Script/Stage1_DeepLabV3_Segmentation.ipynb Script/Stage2_Tumor_Specific_Features.ipynb Script/Stage(3_4_5)_Deep Features_Features Fusion_Classification.ipynb Script/Stage6_Dual-Channel Visual–Textual Explainability.ipynb

📜 Citation

If you use XMorph in your research, please cite our work

📂 Repository Structure

.
├── Script/                              # Sequential execution notebooks
│   ├── Stage1_DeepLabV3_Segmentation.ipynb
│   ├── Stage2_Tumor_Specific_Features.ipynb
│   ├── Stage(3_4_5)_Deep Features_Features Fusion_Classification.ipynb
│   └── Stage6_Dual-Channel Visual–Textual Explainability.ipynb
├── src/                                 # Source data and assets
│   ├── Dataset/                         # CE-T1 MRI samples organized by class
│   ├── figure/                          # Result plots (ROC, Grad-CAM, etc.)
│   ├── labels.npy                       # Ground truth class labels
│   ├── llm_prompts_testset.csv          # Structured data for reproducible GPT-5 inference
│   └── logo.png                         # Project branding
├── requirements.txt                     # Python dependencies
└── README.md                            # This file
---