⚡ electrical engineering

DiffSegLung: Diffusion Radiomic Distillation for Unsupervised Lung Pathology Segmentation

DiffSegLung is an unsupervised framework for lung pathology segmentation that leverages Diffusion Radiomic Distillation, using handcrafted radiomic descriptors as a physics-grounded teacher to guide a 3D diffusion U-Net, thereby achieving superior performance across multiple pathology classes without requiring annotated training data.

Original authors: Rezkellah Noureddine Khiati, Pierre-Yves Brillet, Catalin Fetita

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rezkellah Noureddine Khiati, Pierre-Yves Brillet, Catalin Fetita

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to spot different types of damage inside a human lung using only X-ray pictures (CT scans). The catch? You have no teacher to show the robot what the damage looks like. You don't have any "answer keys" or labeled examples. This is the "unsupervised" problem the paper tackles.

Here is how the authors, Rezkellah Noureddine KHIATI and colleagues, solved it with their new system, DiffSegLung.

The Problem: The Robot is Blind to Physics

Usually, when computers look at medical scans, they treat the image like a regular photo. They look at colors and shapes. But in a CT scan, the "color" (actually a number called a Hounsfield Unit, or HU) tells you exactly how dense the tissue is.

Healthy lung: Very airy (low density).
Emphysema: Even airier (very low density).
Fibrosis (scarring): Dense and heavy (high density).

Previous AI models were like a student who only looked at the picture but ignored the numbers. They treated the image as a simple 8-bit photo, throwing away the crucial density information that actually tells a doctor what's wrong.

The Solution: A "Physics Teacher" and a "Student"

The authors created a system with two main characters: a Student and a Teacher.

The Student (The AI): This is a 3D "diffusion model." Think of it as an artist who is very good at recreating images from scratch. As it learns to redraw the lung scans, it builds a mental map (a "bottleneck") of what the lung looks like.
The Teacher (The Physics Guide): Since they have no human labels, they can't use a human teacher. Instead, they use a set of handcrafted math rules (called "radiomic descriptors").
- Analogy: Imagine the Teacher is a strict librarian who doesn't know what a "book" is, but knows exactly how to measure the thickness of the paper, the texture of the cover, and the weight of the book.
- The Teacher measures these physical properties of the lung tissue. It doesn't care about the "picture"; it cares about the physics.

The Training: "Distillation" (Squeezing Knowledge In)

The system uses a process called Radiomic Distillation.

The Student tries to recreate the lung image.
At the same time, the Teacher measures the physical properties of the patch of lung the Student is looking at.
The system forces the Student's internal map to match the Teacher's measurements. If the Student thinks two patches are similar, but the Teacher says their physical textures are totally different, the Student gets "corrected."
The Magic: This happens without the system ever seeing a single labeled example of "emphysema" or "fibrosis." It just learns to organize the lung data based on physical reality.

The Inference: Sorting the Clutter

Once the Student is trained, the Teacher is thrown away (like a training wheel). Now, the system has to do the actual segmentation (drawing the boundaries of the diseases).

The GMM Sorter: The system looks at all the different patches of the lung it has learned about. It uses a statistical method (Gaussian Mixture Model) to group them into clusters, like sorting a pile of mixed-up socks into piles of "white," "black," and "striped."
The HU Labeler: Here is where the system gets smart. It looks at the average "density" (HU) of each pile.
- "This pile is very light? That must be Emphysema."
- "This pile is heavy? That must be Fibrosis."
- It assigns the disease names based on known physics rules, not by guessing.
The Edge Polisher: The initial sorting might look a bit blocky or jagged. The system uses a final step called Sobel-Diffusion Fusion to smooth out the edges, making the boundaries look sharp and natural, like a high-quality sketch.

The Results: Better than the Rest

The team tested this on 190 expert-annotated lung slices (which they kept hidden until the very end to test the AI).

The Score: Their method (DiffSegLung) beat all other unsupervised methods. It was much better at finding the four types of lung diseases (Emphysema, Fibrosis, Consolidation, and Cavitary lesions) than previous AI models that didn't use this "Physics Teacher."
The Bonus: Because the system learned to respect the physical density of the tissue, it also became better at generating (creating) realistic lung images. It didn't just learn to find diseases; it learned to understand the lung better overall.

The One Weakness

The paper admits that Emphysema is still the hardest to spot.

Why? Because Emphysema and healthy lung tissue are both very airy. They sit in a very similar "density range." It's like trying to tell the difference between two shades of white paint that are almost identical. Even with the Physics Teacher, the AI sometimes struggles to draw the line perfectly between healthy air and diseased air.

Summary

DiffSegLung is a way to teach an AI to find lung diseases without showing it any examples. It does this by forcing the AI to learn the physics of the lung (density and texture) using a mathematical "Teacher," and then using those physical rules to sort the lung into healthy and sick areas. It's like teaching a child to sort rocks not by showing them a picture of a "granite" rock, but by teaching them how to feel the weight and texture of the rock until they can sort them by touch alone.

Technical Summary: DiffSegLung

Problem Statement
Unsupervised segmentation of pulmonary pathologies (emphysema, fibrosis, consolidation, and cavitary lesions) in CT scans remains a significant challenge. This difficulty arises from the scarcity of annotated multi-pathology cohorts and the severe class imbalance inherent in medical data, which renders supervised and semi-supervised approaches impractical. Furthermore, existing diffusion-based segmentation methods fail to exploit the quantitative Hounsfield Unit (HU) signal, which physically distinguishes tissue classes. Prior CT diffusion models typically train on 8-bit normalized images, discarding critical quantitative information, and do not incorporate handcrafted radiomic descriptors to guide representation learning.

Methodology: DiffSegLung
The authors propose DiffSegLung, a framework for fully unsupervised lung pathology segmentation that introduces Diffusion Radiomic Distillation. The framework operates in two stages:

Training Phase (Diffusion-Radiomic Distillation):
- Architecture: A 3D pixel-space Denoising Diffusion Probabilistic Model (DDPM) with a U-Net backbone is trained on native, HU-preserved 3D patches.
- Teacher-Student Distillation: Instead of a neural network teacher, the system uses a set of handcrafted, non-differentiable radiomic descriptors as a "physics-grounded teacher." For each patch, a 34-dimensional radiomic vector is extracted, comprising GLCM statistics, LBP histograms, Gabor filter responses, and first-order HU statistics.
- Contrastive Objective: A radiomic projection head ( $\phi_r$ ) maps the teacher's radiomic vector to a shared embedding space, while a student projection head ( $\phi_s$ ) maps the U-Net bottleneck features to the same space. An InfoNCE contrastive loss aligns these projections: patch pairs with high radiomic cosine similarity are pulled together, while dissimilar pairs are pushed apart.
- Warmup Schedule: To prevent the bottleneck representation from collapsing before meaningful texture features are established, the distillation loss is introduced gradually via a warmup schedule.
- Data Preservation: The model is trained on native HU values rather than normalized 8-bit images to preserve physical density information.
Inference Phase:
- Feature Extraction: The trained encoder ( $\epsilon_\theta$ ) is run using a DPM-Solver for 250 steps. Bottleneck features are recorded at multiple timesteps ( $T = \{50, 100, 150, 200\}$ ).
- Clustering: Multi-timestep features are aggregated and passed through the student projection head. These embeddings are clustered using a Gaussian Mixture Model (GMM) with 5 components.
- Label Assignment: Cluster labels are assigned based on the mean HU value of each cluster compared against clinically established thresholds. This is the only point where domain knowledge is applied; no labeled data is used.
- Boundary Refinement: A Sobel-Diffusion Fusion module refines the segmentation boundaries by combining gradient-based edge detection (Sobel) with the cluster decision surface, followed by a physics-aware HU compatibility check to suppress predictions falling outside expected HU ranges for the assigned class.

Key Contributions
The paper outlines three primary contributions:

Radiomic Teacher-Student Distillation: The first scheme to use handcrafted radiomic descriptors as a teacher for diffusion representation learning, requiring no manual annotations.
HU-Preserving Training: A 3D DDPM training strategy that retains quantitative HU information, producing physically meaningful bottleneck representations.
Improved Performance: Demonstrated improvements in both unsupervised segmentation coherence and generation fidelity compared to prior baselines.

Experimental Results
The framework was evaluated on 190 expert-annotated axial slices drawn from four heterogeneous CT cohorts (BPCO, ILD, ILD-2, and Tuberculosis).

Segmentation Performance: DiffSegLung outperformed all unsupervised baselines (K-Means on radiomics, Diffusion features only, and DAAM) across all four pathology classes.
- Overall DSC: Achieved 89.3% (mean over four classes), compared to 81.6% for the next best baseline (DAAM).
- Class-Specific Performance: The model achieved DSC scores of 84.1% (Healthy), 86.4% (GGO), 76.2% (Fibrosis), and 8.6% (Emphysema). The authors note that emphysema remains the most challenging class due to its HU overlap with healthy parenchyma ($-950$ to $-700$ HU), which limits separability even with radiomic distillation.
- Boundary Accuracy: The method achieved a mean HD95 of 8.6, significantly lower (better) than baselines, attributed to the Sobel-Diffusion Fusion refinement.
Generation Quality: DiffSegLung improved generation fidelity over prior CT diffusion models (LungDDPM+ and MedVAE).
- FID: Reduced from 31.7 (LungDDPM+) to 18.4.
- SSIM/PSNR: Improved to 0.891 and 31.8, respectively.
- Ablation Study: Confirmed that HU preservation provided the largest single gain (+3.5 DSC points). The warmup schedule was critical to prevent bottleneck collapse, and multi-timestep aggregation improved boundary localization.

Significance and Claims
The authors claim that DiffSegLung demonstrates that radiomic distillation can benefit both discriminative (segmentation) and generative objectives simultaneously. By using handcrafted physical descriptors as a teacher, the framework imposes a physically consistent structure on the latent space of a generative model without requiring annotated data. The work highlights that preserving quantitative HU information is the single most important component for distinguishing tissue classes, particularly for pathologies like emphysema where density is the primary differentiator. The paper concludes that while the method significantly advances unsupervised segmentation, the inherent HU overlap between emphysema and healthy tissue remains a limitation, suggesting future work should focus on finer-grained HU-aware clustering.