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

Source: arXiv - 2602.21178v1

Overview

The paper introduces XMorph, a new framework that blends deep‑learning image analysis with large language models (LLMs) to deliver both high‑accuracy brain‑tumor classification and human‑readable explanations. By focusing on tumor boundaries and coupling visual heatmaps with AI‑generated text, the authors aim to bridge the gap between cutting‑edge research and real‑world clinical deployment.

Key Contributions

• Information‑Weighted Boundary Normalization (IWBN): A novel preprocessing step that amplifies diagnostically important boundary pixels, improving the model’s sensitivity to irregular tumor margins.
• Hybrid Morphological Feature Set: Combines traditional deep features with chaotic‑system descriptors and clinically validated shape metrics for richer tumor representation.
• Dual‑Channel Explainable AI Module: Merges Grad‑CAM++ visual explanations with LLM‑produced textual rationales, turning opaque predictions into clinician‑friendly narratives.
• High Performance with Low Compute: Achieves 96 % classification accuracy on the three‑class brain‑tumor benchmark while keeping inference time and memory footprint suitable for edge or hospital‑grade hardware.
• Open‑Source Release: Full code, pretrained weights, and a demo notebook are publicly available, encouraging reproducibility and community extensions.

Methodology

1. Data Preparation – Standard MRI slices of glioma, meningioma, and pituitary tumors are pre‑processed (bias field correction, intensity normalization).
2. IWBN Layer – The image passes through an “information‑weighted” mask that highlights pixels near tumor edges based on a learned boundary importance map. This forces the downstream CNN to focus on shape irregularities rather than just texture.
3. Feature Extraction
   • A lightweight CNN backbone extracts deep visual features.
   • In parallel, the system computes chaotic‑system metrics (e.g., Lyapunov exponents) and classic morphological descriptors (area, perimeter, compactness).
   • These are concatenated into a hybrid feature vector.
4. Classification Head – A fully‑connected classifier predicts one of the three tumor types.
5. Explainability Pipeline
   • Grad‑CAM++ produces a heatmap that visualizes which image regions contributed most to the decision.
   • The heatmap and the hybrid feature vector are fed to a fine‑tuned LLM (e.g., LLaMA‑2) that generates a concise textual rationale (e.g., “The lesion shows an irregular, spiculated border with high chaotic entropy, typical of glioma”).
6. Training – End‑to‑end training uses cross‑entropy loss plus a boundary‑regularization term that encourages the IWBN mask to stay focused on true tumor edges.

Results & Findings

• Boundary Emphasis Pays Off: Ablation studies show that removing IWBN drops overall accuracy by ~3 %, confirming the importance of edge‑focused learning.
• Explainability Without Sacrifice: Adding the LLM rationale incurs <5 ms overhead and does not degrade classification performance.
• Robustness: The model maintains >90 % accuracy when tested on unseen MRI scanners, indicating good generalization.

Practical Implications

• Clinical Decision Support: Radiologists can receive a quick prediction and an AI‑generated explanation that mirrors the language used in tumor boards, facilitating trust and faster verification.
• Edge Deployment: The lightweight backbone and efficient IWBN make it feasible to run on hospital PACS servers or even on portable MRI units, expanding access in low‑resource settings.
• Regulatory Readiness: Transparent visual and textual explanations align with emerging AI‑medical device guidelines that demand “human‑interpretable” outputs.
• Research Extension: The open‑source codebase allows developers to plug in other LLMs (e.g., domain‑specific medical LLMs) or adapt the boundary‑normalization concept to other segmentation‑heavy tasks such as lung nodule detection or retinal disease grading.

Limitations & Future Work

• Dataset Scope: Experiments are limited to a publicly available three‑class MRI dataset; broader multi‑institutional validation is needed to confirm clinical robustness.
• LLM Hallucination Risk: While the generated rationales are generally accurate, occasional mismatches between text and heatmap were observed, suggesting the need for tighter grounding mechanisms.
• Boundary Ground Truth: IWBN relies on implicit edge learning; explicit boundary annotations could further improve performance but would increase labeling cost.
• Future Directions: The authors plan to explore self‑supervised pretraining on large unlabeled MRI corpora, integrate multimodal data (e.g., patient history), and evaluate the system in prospective clinical trials.

Authors

• Sepehr Salem Ghahfarokhi
• M. Moein Esfahani
• Raj Sunderraman
• Vince Calhoun
• Mohammed Alser

Paper Information

• arXiv ID: 2602.21178v1
• Categories: cs.CV, cs.AI
• Published: February 24, 2026
• PDF: Download PDF