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

XMorph is an explainable, computationally efficient framework that combines an Information-Weighted Boundary Normalization mechanism with a dual-channel LLM-assisted AI module to achieve 96.0% accuracy in classifying glioma, meningioma, and pituitary tumors while providing clinically interpretable visual and textual insights.

The Gist

Imagine you are a detective trying to solve a mystery inside a patient's brain. The clues are hidden in MRI scans, and the suspects are three types of brain tumors: Gliomas (the aggressive, messy ones), Meningiomas (the lumpy but usually slow-growing ones), and Pituitary tumors (the smooth, contained ones).

For a long time, computers have been good at looking at these scans and saying, "I think it's a Glioma!" But there was a big problem: the computer was a "Black Box." It gave the answer, but it couldn't explain why. It was like a magic 8-ball that just said "Yes" or "No" without showing its work. Doctors didn't trust it because they couldn't see the reasoning.

Enter XMorph, a new AI system created by researchers at Georgia State University. Think of XMorph not as a magic 8-ball, but as a super-smart medical detective with a clear notebook and a translator.

Here is how XMorph solves the mystery, broken down into simple steps:

1. The "Smart Outline" (Segmentation)

First, the computer needs to find the tumor. It's like trying to find a specific shape in a messy drawing. XMorph uses a tool called DeepLabV3 to draw a perfect, tight outline around the tumor, separating it from the healthy brain tissue. It's so good at this that it rarely mistakes a healthy spot for a tumor or misses a tiny part of the tumor.

2. The "Shape Detective" (The New Secret Sauce)

This is where XMorph gets really clever. Most AI just looks at the texture of the tumor (like looking at the color of a wall). But XMorph also looks at the shape of the edge.

• The Analogy: Imagine the tumor's edge is a coastline.
  • A Pituitary tumor is like a smooth, round lake.
  • A Meningioma is like a coastline with some bays and inlets (lobes).
  • A Glioma is like a jagged, chaotic cliffside with deep, irregular cracks.

XMorph uses a new trick called IWBN (Information-Weighted Boundary Normalization). Imagine giving a magnifying glass to the computer, but the magnifying glass automatically zooms in only on the jagged, messy parts of the edge and ignores the smooth parts. It turns that jagged coastline into a mathematical "sound wave" to measure how chaotic it is. This helps it spot the dangerous, infiltrative tumors that other AI might miss.

3. The "Two-Brain" Team (Hybrid Intelligence)

XMorph doesn't rely on just one way of thinking. It uses a dual-brain approach:

• Brain A (The Deep Learner): This is a standard AI that looks at the whole picture and learns from millions of other images. It's great at recognizing patterns.
• Brain B (The Math Detective): This is the part that measures the jagged edges, the chaos, and specific medical signs (like how much the tumor is pushing the brain to one side).

XMorph combines the "gut feeling" of Brain A with the "hard math" of Brain B. This makes the final decision much stronger and more accurate than using either one alone.

4. The "Translator" (Explainable AI)

This is the most important part for doctors. When XMorph makes a diagnosis, it doesn't just spit out a result. It uses two channels to explain itself:

• Channel 1: The Heatmap (Visual): It highlights the exact spots on the MRI that made it suspicious, like a detective circling the crime scene on a map.
• Channel 2: The Translator (Text): This is where a Large Language Model (LLM) steps in. It takes the complex math numbers (like "high chaos score" or "jagged edge") and translates them into plain English.

Example: Instead of saying "Feature 45 is 0.9," XMorph says:

"I believe this is a Glioma because the tumor has a very jagged, chaotic edge and is pushing the brain's center line significantly. These are classic signs of an aggressive, infiltrative tumor."

Why Does This Matter?

• Trust: Doctors can finally see why the computer made a choice. It's no longer a black box; it's a partner with a clear explanation.
• Speed: It's fast and doesn't need a supercomputer to run, making it usable in real hospitals.
• Accuracy: It got 96% accuracy in tests, correctly identifying the three main types of brain tumors.

The Bottom Line

XMorph is like upgrading from a robot that just guesses to a consultant who can show you the evidence and explain the logic. By combining advanced math (to measure the tumor's shape) with modern language AI (to explain it to humans), it bridges the gap between high-tech computing and real-world medical care. It proves that AI can be both incredibly smart and perfectly understandable.

Technical Summary

1. Problem Statement

The paper addresses three critical barriers hindering the clinical adoption of AI for brain tumor diagnosis:

• The "Black Box" Problem: Deep learning models (e.g., CNNs) lack transparency, making it difficult for clinicians to trust predictions without understanding the reasoning.
• Computational Constraints: High-performing models are often computationally expensive, limiting their use in real-time or resource-constrained clinical settings (e.g., portable MRI).
• Inadequate Morphological Modeling: Conventional models often fail to capture the complex, irregular, and non-Euclidean boundaries characteristic of malignant tumor growth. They rely heavily on visual textures but miss subtle, clinically relevant boundary irregularities that can be quantified using nonlinear dynamics.

2. Methodology: The XMorph Framework

XMorph is a six-stage, hybrid pipeline that integrates deep learning, nonlinear dynamics, and Large Language Models (LLMs) to classify glioma, meningioma, and pituitary tumors.

Stage 1: Automated Tumor Segmentation

• Model: DeepLabV3 with a ResNet-50 backbone.
• Optimization: Uses a hybrid loss function (Cross-Entropy + Dice Loss) to handle class imbalance and ensure precise boundary delineation.
• Output: A binary mask isolating the tumor Region of Interest (ROI).

Stage 2: Tumor-Specific Feature Extraction (The Core Innovation)

This stage converts the 2D tumor boundary into a 1D radial signal to extract interpretable features:

• Boundary-to-Signal Conversion: The tumor contour is resampled to 256 points and normalized by mean radius to create a scale-invariant time series.
• Geometric & Statistical Features: Calculates Irregularity Index (global fluctuation) and Roughness Index (fine-scale jaggedness).
• Information-Weighted Boundary Normalization (IWBN): A novel mechanism that computes local entropy at each boundary point. It generates weights to amplify regions of high structural complexity, creating a weighted signal that highlights diagnostically significant irregularities.
• Nonlinear Chaotic Descriptors: Extracts Fractal Dimension (FD), Approximate/Sample/Permutation Entropy, and the Largest Lyapunov Exponent (LE) to quantify chaos and self-organization in tumor growth.
• Clinical Biomarkers: Integrates radiological metrics: Ring Enhancement Index (REI), Midline Shift (MLS), and Skull-to-Tumor Distance.

Stage 3: Deep Feature Extraction

• Model: Pre-trained ResNet-50 processes the full MRI slice (not just the ROI) to capture high-level textural and contextual features.
• Processing: Features are extracted from the global average pooling layer and reduced via Principal Component Analysis (PCA) to maintain efficiency.

Stage 4: Hybrid Feature Fusion

• Concatenates the Tumor-Specific Features (Stage 2) and Deep Features (Stage 3) into a unified vector. This combines the interpretability of handcrafted morphological/clinical features with the semantic power of deep learning embeddings.

Stage 5: Classification

• Classifier: XGBoost (Gradient Boosting).
• Advantage: Chosen for its computational efficiency and ability to handle the fused feature vector effectively without the heavy inference cost of deep neural networks.

Stage 6: Dual-Channel Explainable AI (XAI)

• Visual Channel: Uses GradCAM++ to generate heatmaps highlighting the specific MRI regions influencing the decision.
• Textual Channel: Uses an LLM (GPT-5) to generate clinical narratives. The LLM receives the top-k SHAP (SHapley Additive exPlanations) values from the XGBoost model, translating quantitative metrics (e.g., "high local entropy," "large midline shift") into human-readable diagnostic rationales.

3. Key Contributions

1. Information-Weighted Boundary Normalization (IWBN): A first-of-its-kind method that dynamically weights boundary segments based on local entropy, amplifying diagnostically relevant irregularities that global metrics miss.
2. Hybrid Feature Representation: A unified framework fusing CNN embeddings, nonlinear chaotic metrics (FD, ApEn, LE), and clinical biomarkers (REI, MLS), providing a richer tumor characterization than either approach alone.
3. Dual-Channel XAI: The first framework to couple spatial visualizations (GradCAM++) with LLM-generated textual rationales, grounding AI predictions in established clinical language and nonlinear dynamics.
4. Efficiency: Demonstrates that high accuracy can be achieved with a lightweight architecture (XGBoost + ResNet-50) compared to heavy, end-to-end deep learning models.

4. Experimental Results

• Dataset: Evaluated on the Figshare 2024 collection (3,564 MRI scans: Glioma, Meningioma, Pituitary, and Non-Tumor).
• Segmentation: DeepLabV3 achieved a Dice Score of 0.932, outperforming U-Net (0.897) and showing superior stability.
• Classification Performance:
  • Tumor-Specific Features Only: 90.0% Accuracy.
  • Deep Features Only: 93.0% Accuracy.
  • Fused Hybrid Features: 96.0% Accuracy (0.962 Sensitivity, 0.983 Specificity).
• Feature Analysis:
  • Gliomas showed the highest Irregularity Index and Mean Local Entropy, confirming the model's ability to detect chaotic, infiltrative growth.
  • Pituitary tumors exhibited smooth, regular boundaries with low entropy.
  • Meningiomas showed intermediate complexity with localized lobulations.
• Explainability: The LLM successfully translated complex SHAP values (e.g., "high boundary entropy") into clinical insights (e.g., "consistent with diffuse, infiltrative pathology"), bridging the gap between mathematical descriptors and radiological reasoning.

5. Significance

• Clinical Trust: By providing both visual heatmaps and natural language explanations, XMorph addresses the "black box" issue, making AI decisions transparent and actionable for radiologists.
• Bridging Disciplines: It uniquely integrates nonlinear dynamics (chaos theory) with deep learning and clinical practice, offering a more holistic view of tumor morphology than standard CNNs.
• Efficiency: The framework proves that lightweight, hybrid models can outperform or match heavy deep learning architectures while being computationally feasible for real-world deployment.
• Future Impact: Sets a precedent for "Hybrid Deep Intelligence" in medical imaging, suggesting a path toward systems that are not only accurate but also mathematically rigorous and clinically interpretable.