Diffusion-Based Feature Denoising and Using NNMF for Robust Brain Tumor Classification
| Authors | Hiba Adil Al-kharsan & Róbert Rajkó |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2603.13182 |
| Download | |
| Categories | cs.CV |
Abstract
Brain tumor classification from magnetic resonance imaging, which is also known as MRI, plays a sensitive role in computer-assisted diagnosis systems. In recent years, deep learning models have achieved high classification accuracy. However, their sensitivity to adversarial perturbations has become an important reliability concern in medical applications. This study suggests a robust brain tumor classification framework that combines Non-Negative Matrix Factorization (NNMF or NMF), lightweight convolutional neural networks (CNNs), and diffusion-based feature purification. Initially, MRI images are preprocessed and converted into a non-negative data matrix, from which compact and interpretable NNMF feature representations are extracted. Statistical metrics, including AUC, Cohen's d, and p-values, are used to rank and choose the most discriminative components. Then, a lightweight CNN classifier is trained directly on the selected feature groups. To improve adversarial robustness, a diffusion-based feature-space purification module is introduced. A forward noise method followed by a learned denoiser network is used before classification. System performance is estimated using both clean accuracy and robust accuracy under powerful adversarial attacks created by AutoAttack. The experimental results show that the proposed framework achieves competitive classification performance while significantly enhancing robustness against adversarial perturbations.The findings presuppose that combining interpretable NNMF-based representations with a lightweight deep approach and diffusion-based defense technique supplies an effective and reliable solution for medical image classification under adversarial conditions.
Engineering Breakdown
Plain English
This paper addresses adversarial robustness in brain tumor classification from MRI scans, a critical problem since deep learning models can be fooled by small perturbations in medical imaging. The authors propose a framework combining Non-Negative Matrix Factorization (NNMF) for feature extraction, lightweight CNNs for classification, and diffusion-based feature purification to defend against adversarial attacks. The approach preprocesses MRI images into non-negative matrices, extracts interpretable NNMF features, and evaluates performance using statistical metrics like AUC, Cohen's d, and p-values. This addresses a major reliability concern: while deep learning achieves high accuracy on brain tumor classification, its vulnerability to adversarial perturbations makes it unsafe for clinical deployment without additional safeguards.
Core Technical Contribution
The core novelty is the integration of NNMF with diffusion-based feature purification as a defense mechanism against adversarial attacks on medical imaging classifiers. Unlike standard CNN approaches that operate directly on pixel values, this framework extracts compact, interpretable features via NNMF and then applies diffusion processes to purify these features before classification, reducing attack surface and improving robustness. The use of NNMF is particularly significant because it enforces non-negativity constraints, which are naturally suited to MRI intensity values and may inherently reduce susceptibility to adversarial perturbations. This combines domain-specific preprocessing (non-negative matrix constraints) with modern robustness techniques (diffusion-based purification) rather than relying solely on standard adversarial training methods.
How It Works
The pipeline begins with MRI image preprocessing and conversion into a non-negative data matrix X ∈ ℝ+^(m×n), which represents pixels or patches with non-negative intensity values. NNMF decomposes this into X ≈ WH where W contains learned basis components and H contains activations, producing compact feature representations that are interpretable and constrained to non-negative values. The extracted NNMF features are then passed through a diffusion-based purification step, which smooths and denoises the feature space to remove adversarial perturbations while preserving task-relevant information. These purified features feed into a lightweight CNN for final tumor classification (likely multi-class: benign, malignant, etc.), with performance evaluated via AUC, Cohen's d effect sizes, and statistical significance testing to ensure robust comparisons across tumor types.
Production Impact
For production medical imaging systems, this framework directly solves the adversarial robustness problem that prevents deep learning models from being clinically deployed without certification. Engineers integrating this would replace standard preprocessing + CNN pipelines with the NNMF + diffusion + lightweight CNN stack, adding computational overhead primarily in the NNMF factorization and diffusion steps but reducing model size via lightweight CNNs. The statistical evaluation (AUC, Cohen's d, p-values) provides clinically meaningful confidence intervals and effect sizes, enabling regulatory compliance and clinical validation workflows that pure accuracy metrics cannot support. Trade-offs include: NNMF is computationally more expensive than direct CNN feature extraction upfront, the diffusion step adds latency (~milliseconds per inference), but the compact representations reduce memory and enable deployment on edge devices (e.g., hospital imaging workstations with limited GPU resources).
Limitations and When Not to Use This
The abstract is truncated, so specific performance numbers, baseline comparisons, and full method details are unavailable, making it difficult to assess whether robustness gains come at the cost of clean accuracy degradation. The approach assumes MRI scans are always non-negative after preprocessing, which may not hold universally (e.g., after intensity normalization or artifact removal), and the framework's transferability to other medical imaging modalities (CT, ultrasound, X-ray) is unclear. The paper does not address query complexity or adaptive attacks where adversaries know about the NNMF + diffusion defense and craft perturbations to circumvent it—standard NNMF without certified robustness guarantees may be vulnerable to sufficiently crafted attacks. Follow-up work should include certified robustness bounds, evaluation on multiple tumor classification datasets, and comparison against other medical imaging robustness techniques (e.g., certified defenses, randomized smoothing).
Research Context
This work builds on two distinct research streams: adversarial robustness in medical AI (responding to papers showing CNNs can be fooled by imperceptible perturbations) and interpretability via dimensionality reduction (NNMF has a long history in medical imaging for parts-based feature discovery). The paper advances beyond standard adversarial training by leveraging domain structure (non-negativity in MRI) and integrating diffusion models, which have recently gained traction as a general robustness mechanism across vision tasks. This likely targets benchmark datasets like BraTS (Brain Tumor Segmentation) or proprietary clinical MRI collections, and positions itself in the intersection of medical computer vision, adversarial robustness, and explainability—three growing areas as regulatory bodies demand AI safety certification in healthcare.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::