💻 computer science

H-SemiS: Hierarchical Fusion of Semi and Self-Supervised Learning for Knee Osteoarthritis Severity Grading

H-SemiS is a hierarchical semi-supervised framework that improves knee osteoarthritis severity grading by decomposing multi-class classification into binary sub-tasks and integrating adversarial self-supervised reconstruction with quantum-inspired feature mixing to better utilize unlabeled data and handle class imbalance.

Original authors: Chandravardhan Singh Raghaw, Anushka Parwal, Shahid Shafi Dar, Prajakta Darade, Nagendra Kumar

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Chandravardhan Singh Raghaw, Anushka Parwal, Shahid Shafi Dar, Prajakta Darade, Nagendra Kumar

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 a librarian in a massive, ancient library. You have millions of books, but there is a huge problem: almost none of them have covers or titles. You know some are about history, some are about science, and some are about poetry, but because they are unlabeled, you can’t easily tell them apart. Even worse, some books are very rare (like a single, precious book on ancient philosophy), while others are everywhere (like common cookbooks).

This is exactly the problem doctors face with Knee Osteoarthritis (KOA). To treat it, doctors need to grade how severe the disease is (from "no issues" to "severe damage"). This requires looking at thousands of X-rays, but having experts sit and label every single one is incredibly expensive and slow.

The researchers created a system called H-SemiS to solve this. Here is how it works, using three simple metaphors:

1. The "Master Artist" (Self-Supervised Learning)

Since most X-rays don't have labels, the system starts by playing a game of "Fill in the Blanks."

Imagine taking a beautiful painting, covering up 75% of it with black squares, and asking an apprentice to redraw the missing parts. To do this well, the apprentice has to truly understand what a knee looks like—where the bones are, how the joints curve, and how the shadows fall.

By practicing this "reconstruction" millions of times on unlabeled X-rays, the AI becomes a "Master Artist." It learns the deep, structural secrets of the human knee without ever needing a doctor to tell it what it's looking at.

2. The "Smart Sorting Hat" (Hierarchical Classification)

Usually, if you ask an AI to pick between five different grades of severity (Grade 0 to Grade 4), it gets overwhelmed—especially if there are way more "Grade 0" images than "Grade 4" images. It’s like asking a child to sort a giant pile of LEGOs into five different colored bins all at once; they’ll likely get confused.

Instead, H-SemiS uses a "Decision Tree" approach. It asks simple "Yes/No" questions:

Question 1: "Is this knee healthy or does it have issues?"
Question 2 (If it has issues): "Is it a mild issue or a serious one?"

By breaking one big, scary task into a series of small, easy choices, the AI stays accurate and doesn't get "distracted" by the fact that some grades are much more common than others.

3. The "Quantum Microscope" (Quantum-Inspired Learning)

Sometimes, the difference between a "mild" knee and a "moderate" knee is incredibly subtle—like trying to tell the difference between two nearly identical shades of gray. In a standard X-ray, these details can get blurred or hidden by overlapping shadows.

The researchers added a "Quantum" twist. Think of this as upgrading from a standard magnifying glass to a magical, high-tech microscope that can see "between the lines." It uses mathematical principles inspired by quantum physics to spot complex, non-linear patterns that a normal computer might miss. It allows the AI to see the "hidden geometry" of the joint, making it much better at distinguishing between two very similar stages of the disease.

The Bottom Line

By combining these three tricks—learning from unlabeled data (The Artist), breaking big tasks into small ones (The Sorting Hat), and seeing hidden patterns (The Quantum Microscope)—the H-SemiS framework can grade knee health almost as accurately as a human expert, even when it has very little "instruction manual" (labeled data) to work with.

This means faster, cheaper, and more accurate early detection for millions of people suffering from joint pain.

Technical Summary: H-SemiS Framework for Knee Osteoarthritis Severity Grading

1. Problem Statement

Knee Osteoarthritis (KOA) is a major degenerative disease requiring accurate severity grading using the Kellgren-Lawrence (KL) system (Grades 0–4). Automated grading via knee X-rays (KXR) faces several critical challenges:

Data Scarcity: High-quality, expert-annotated medical datasets are expensive and difficult to obtain.
Class Imbalance: Radiographic samples are often skewed toward specific severity levels, making standard multi-class classification difficult.
Image Complexity: Diagnostic features like Joint Space Narrowing (JSN) are often obscured by low contrast, blurred pixels, overlapping anatomical structures, and varying illumination.
Label Noise: Semi-supervised methods often rely on "pseudo-labels" that can be noisy or ambiguous, leading to model instability.

2. Proposed Methodology (H-SemiS)

The authors propose H-SemiS (Hierarchical fusion of Semi and Self-Supervised learning), a three-stage framework designed to maximize learning from limited labeled data and unlabeled datasets.

Stage 1: Masked Image Reconstruction (MI-Rec)
To leverage unlabeled data, the framework employs a self-supervised module inspired by Masked Autoencoders (MAE) and Generative Adversarial Networks (GAN).

Process: 75% of image patches are randomly masked. A Reconstruction Generator attempts to rebuild the missing regions, while a Reconstruction Discriminator ensures the reconstructed patches are anatomically realistic.
Purpose: This forces the model to learn robust anatomical structures and generates "proxy data" to augment the training set.

Stage 2: Similarity-aware Reconstructed Image Labeler (SiRL)
Since the reconstructed images from Stage 1 are unlabeled, SiRL assigns them proxy labels.

Process: The system constructs a "template library" of feature vectors representing each KL grade. It calculates a combined score using Cosine Similarity and Euclidean Distance between the reconstructed sample and the templates.
Filtering: A threshold ( $\tau = 0.80$ ) is applied; samples with low similarity are discarded to prevent noisy labels from corrupting the training process.

Stage 3: Hierarchical Quantum-infused Teacher-Student (HQ-TeSt)
This stage performs the actual classification using a hybrid classical-quantum approach.

Quantum-infused Teacher-Student (Q-TeSt): A Mean-Teacher architecture where a "Student" model learns from labeled/proxy data, and a "Teacher" model provides consistent targets for unlabeled data via Exponential Moving Average (EMA).
Quantum Convolutional Network (QCN): To capture nonlinear, complex anatomical patterns, the framework maps classical features into a Hilbert space using amplitude encoding. It utilizes parameterized quantum gates (rotation and CNOT) to exploit quantum entanglement and superposition for better feature discrimination.
Hierarchical Multi-classifier (HiM): Instead of a single 5-class output, the task is decomposed into a binary decision tree (e.g., Normal vs. Osteoarthritis $\rightarrow$ Grade 2 vs. Grade 3, etc.). This directly addresses class imbalance by ensuring each decision node handles balanced binary subsets.

3. Key Contributions

Novel Fusion: The first framework to integrate adversarial self-supervision with semi-supervised teacher-student learning specifically for KOA.
Quantum Integration: The pioneering use of quantum principles (QCN) to resolve complications arising from overlapping anatomical structures in X-rays.
Imbalance Mitigation: A dual approach using both hierarchical decomposition and the generation of reconstructed proxy samples to balance the dataset.
Noise Reduction: The SiRL module provides a robust mechanism to filter out unreliable pseudo-labels.

4. Experimental Results

The framework was evaluated on the OAI and DKXI datasets, comparing it against 13 baselines (supervised, self-supervised, and semi-supervised).

Superior Accuracy: H-SemiS consistently outperformed state-of-the-art models. On the OAI dataset, it achieved an Accuracy of 84.9% and an F1-score of 84.7%, surpassing established methods like Mean-Teacher and various Transformer-based models.
Efficiency: Despite the complexity of quantum simulation, the model maintained a highly competitive inference time (4.09 ms) and balanced parameter count.
Interpretability: XAI evaluations (Grad-CAM, Score-CAM) and the ADCC score confirmed that the model's "attention" aligns with clinically relevant regions (like the joint space).
Generalizability: Cross-dataset testing (training on OAI, testing on DKXI) proved the model is robust to different imaging distributions.

5. Significance

H-SemiS represents a significant advancement in medical AI by demonstrating that quantum-inspired deep learning and hierarchical decomposition can overcome the traditional limitations of medical imaging: small datasets and class imbalance. Its ability to function effectively with only 20% labeled data makes it a highly practical solution for clinical environments where expert annotation is a bottleneck.