💻 computer science

HLNet: High-Confidence Lightweight Network for Semi-Supervised Medical Image Segmentation

This paper proposes HLNet, a high-confidence lightweight network for semi-supervised medical image segmentation that utilizes dual-level perturbations and Evidence Inference Learning (EVIL) to mitigate erroneous pseudo-label supervision and reduce model complexity, achieving superior performance on public datasets with only 20% labeled data.

Original authors: Xiaoyu Huang, Jianxin Hou, Canfei Xu, Jinwei Ge, San Li, Yang Li

Published 2026-09-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaoyu Huang, Jianxin Hou, Canfei Xu, Jinwei Ge, San Li, Yang Li

Original paper licensed under CC BY 4.0 (https://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

In the world of modern medicine, a doctor's ability to diagnose and treat often depends on their capacity to see inside the human body with perfect clarity. Medical images, such as MRI scans or ultrasound pictures, are the windows through which physicians view organs, tissues, and tumors. To make sense of these complex pictures, computers are increasingly used to automatically trace the outlines of specific areas, a process known as segmentation. This digital outlining helps doctors measure the size of a tumor or the volume of a heart chamber with a speed and consistency that human eyes alone cannot match. However, teaching a computer to do this work has traditionally required a massive amount of help from human experts. For a computer to learn how to recognize a heart muscle or a breast lesion, it must be shown thousands of images where a doctor has painstakingly drawn the exact boundaries by hand. This manual labeling is slow, expensive, and often impossible to obtain in sufficient quantities for every type of medical condition.

To solve this shortage of labeled data, scientists have turned to a strategy called semi-supervised learning. This approach allows a computer to learn from a small number of expertly labeled images while also studying a much larger pool of unlabeled images that the computer has not yet seen. The computer tries to guess the boundaries in the unlabeled images and uses those guesses to improve itself, much like a student who learns from a few textbooks and then practices on a vast library of unlabeled notes. Yet, this method has a significant flaw: if the computer makes a mistake early on, it can reinforce that error, teaching itself to be confidently wrong. Furthermore, the powerful computer programs currently used for this task are often too large and heavy to run on portable medical devices, such as the ultrasound machines found at a patient's bedside. Researchers are now seeking a way to build a system that is both smart enough to learn from limited data and light enough to travel with the doctor.

A team of researchers has developed a new solution to these challenges, creating a system they call HLNet. This is a lightweight network designed to be highly confident in its predictions, even when it is working with very little labeled information. The core idea behind their work is to teach the computer to be skeptical of its own guesses until it is sure they are right. Instead of blindly accepting every guess the computer makes about an unlabeled image, the system includes a mechanism to estimate how uncertain it is about a specific pixel. If the computer is unsure, the system ignores that part of the image. If the computer is confident, the system uses that information to guide its learning. This prevents the model from learning from its own mistakes, a common pitfall in previous methods.

To make this system work efficiently on smaller devices, the researchers redesigned the computer's "brain," or backbone. Traditional systems use standard building blocks that process every image in the same rigid way, regardless of what the image actually shows. The new design uses a dynamic component that adjusts its internal settings for each specific image it sees. Think of it as a camera lens that automatically changes its focus and shape to perfectly suit the subject in front of it, rather than using a fixed lens for every photo. This allows the system to capture the unique details of different medical scans without needing a massive amount of memory or processing power. By replacing heavy, standard components with these adaptable, lightweight ones, the researchers created a model that is significantly smaller than its competitors but still capable of high performance.

The researchers tested their new system on three different types of medical data: heart MRI scans, breast ultrasound images, and prostate MRI scans. In these experiments, they gave the computer only a small fraction of the images with expert labels—specifically, just 20% of the data was labeled, while the rest was left unlabeled. Despite this scarcity of guidance, the new system outperformed other advanced methods that had access to the same limited data. It achieved high accuracy in identifying the heart chambers, breast tumors, and prostate tissues, matching the performance of much larger, more complex models that required far more computing power. In fact, on the heart MRI dataset, the new system achieved a level of accuracy that was nearly as good as a fully supervised model trained on 100% of the labeled data, but it did so using only a fifth of the labeled images and a fraction of the computational resources.

A key part of the system's success lies in how it handles the "noise" of uncertainty. When the computer looks at an unlabeled image, it generates a prediction and a measure of how sure it is. The researchers found that by filtering out the pixels where the computer was unsure, they could prevent the model from being misled by bad guesses. They also introduced a self-correction step where the system reviews its own work and refines the final answer before accepting it as a learning signal. This double-checking process ensures that the model only learns from the most reliable information. The results showed that this approach not only improved the accuracy of the segmentation but also made the boundaries of the identified regions much sharper and more precise, which is critical for medical diagnosis.

The study demonstrates that it is possible to build medical image analysis tools that are both highly accurate and practical for real-world use. By combining a lightweight, adaptable architecture with a smart strategy for handling uncertainty, the researchers have created a model that can learn effectively from limited data. This means that in the future, doctors could potentially use advanced segmentation tools on portable devices without needing a supercomputer or a massive database of pre-labeled images. The system proved its worth across different types of medical scans, showing that it can adapt to the unique challenges of imaging the heart, the breast, and the prostate. While the technology is still in the research phase, the findings suggest a path toward more accessible and efficient medical imaging tools that can assist doctors in making faster and more accurate decisions at the point of care.

Technical Summary: HLNet for Semi-Supervised Medical Image Segmentation

Problem Statement

Medical image segmentation is critical for lesion localization and treatment planning, yet it relies heavily on densely labeled data, which is costly and time-consuming to acquire from expert doctors. While Semi-Supervised Learning (SSL) offers a solution by leveraging unlabeled data, current approaches face significant limitations:

Reliance on Low-Quality Pseudo-labels: Most SSL methods use self-training or consistency regularization. These are highly susceptible to error propagation, where low-quality pseudo-labels provide erroneous supervision, degrading model performance.
Inadequate Perturbation Strategies: Existing consistency-based methods often rely on single constraints or single-dimensional perturbations, leading to imprecise decision boundaries and vulnerability to noise.
Computational Burden: State-of-the-art semi-supervised backbones (e.g., CNNs, Vision Transformers) typically have large parameter sizes and high computational costs, making them unsuitable for deployment on portable devices with limited resources, such as bedside ultrasound systems.

Methodology: HLNet Framework

The authors propose HLNet (High-Confidence Lightweight Network), a semi-supervised segmentation model designed to operate under dual-level perturbations while maintaining a lightweight architecture. The framework integrates four core components:

1. Lightweight Backbone with DyBAC

To address computational constraints, HLNet replaces standard convolutions in a UNet variant with the Dynamic Convolution-based Bidirectional Attention Component (DyBAC).

Mechanism: DyBAC integrates a spatial attention mechanism with dynamic convolutions. It decouples spatial and channel correlations using depthwise separable convolutions.
Adaptability: Unlike standard convolutions that use fixed parameters, DyBAC dynamically adjusts convolutional kernel parameters for each input sample based on learned attention coefficients.
Efficiency: This design significantly reduces the parameter count (approximately 1/5th of the baseline Model2) while enhancing feature representation capabilities.

2. Dual-Level Perturbation Strategy

HLNet employs perturbations at two distinct levels to improve robustness and consistency:

Data-Level Perturbation (BCP): The model utilizes Bidirectional Copy-Paste (BCP). In a training batch, labeled and unlabeled images are randomly cropped and pasted onto each other to generate mixed "inward" and "outward" images. This strategy encourages the model to learn from key features and semantic information across data distributions, alleviating distribution mismatch.
Model-Level Perturbation: The architecture employs a shared encoder with two distinct decoders using different upsampling methods. This structural difference acts as a model-level perturbation, forcing the network to learn consistent representations despite architectural variations.

3. Uncertainty Estimation via EVIL

To mitigate the risk of erroneous supervision from low-quality pseudo-labels, HLNet introduces EVidence Inference Learning (EVIL).

Mechanism: One branch of the network acts as an evidential branch, computing uncertainty estimates for pseudo-labels in a single forward pass using evidence theory (Dirichlet distribution).
Filtering: Pixels with high uncertainty are identified and ignored. Only high-confidence pseudo-labels are retained for consistency supervision, effectively filtering out noise.

4. Self-Correction Module (SCM)

To further refine predictions and reduce cognitive uncertainty, a Self-Correction Module (SCM) is added after the student network.

Function: The SCM concatenates predictions from both branches and processes them through a lightweight convolutional block (3x3 conv, BN, ReLU, 1x1 conv).
Goal: This module corrects potential inconsistencies, producing refined predictions that are then supervised by the reliable pseudo-labels.

Loss Function

The total training loss combines three components:
$L_{total} = L_{mix} + L_{mix\_re} + \beta L_{con}$
Where $L_{mix}$ is the mixed loss from BCP, $L_{mix\_re}$ is the corrected mixed loss after SCM, and $L_{con}$ is the mutual consistency loss between branches, weighted by a Gaussian warming function.

Key Contributions

HLNet Architecture: A lightweight semi-supervised network capable of learning from unlabeled data under dual-level perturbations, enhancing generalization and consistency.
DyBAC Integration: The design of a lightweight backbone using DyBAC to replace conventional convolutions, enabling efficient deployment on resource-constrained devices while maintaining accuracy.
Reliability Mechanism: The introduction of EVIL for single-pass uncertainty estimation and the SCM for self-correction, which collectively filter low-quality pseudo-labels to prevent error propagation.
Comprehensive Validation: Extensive experiments demonstrating superior performance across three diverse medical imaging datasets.

Experimental Results

The model was evaluated on three public datasets: ACDC (Cardiac MRI), BUSI (Breast Ultrasound), and Promise12 (Prostate MRI).

Performance vs. Complexity: With only 20% labeled data, HLNet achieved competitive or superior results compared to fully supervised models and other semi-supervised methods (e.g., UAMT, MCNet+, BCP, ABD) while significantly reducing parameters (0.59M) and GFlops (0.87).
ACDC Dataset: HLNet achieved a Dice Similarity Coefficient (DSC) of 89.84% and Jaccard index of 82.08% with 20% labeled data. This outperformed SwinUNet (89.28% DSC) and was comparable to a fully supervised UNet trained on 100% data (91.13% DSC), but with drastically lower computational cost.
BUSI Dataset: The model achieved a DSC of 77.31% with 20% labeled data, outperforming most comparative semi-supervised methods and demonstrating robustness in handling boundary segmentation despite lower image quality.
Promise12 Dataset: HLNet achieved the best or second-best results across all metrics (DSC: 86.08%, Jac: 75.99%) with 20% labeled data, showing strong robustness in prostate segmentation.
Ablation Studies: Experiments confirmed that the combination of BCP, EVIL, and SCM modules provided the most significant performance gains. Specifically, BCP contributed the largest improvement (5.82% DSC increase), while EVIL and SCM further refined the model's confidence and boundary accuracy.

Significance and Claims

The paper claims that HLNet successfully addresses the trade-off between segmentation accuracy, model complexity, and data efficiency in medical imaging. By combining dual-level perturbations with evidence-based uncertainty filtering, the model effectively learns from unlabeled data without succumbing to the error propagation common in self-training methods.

The authors emphasize that HLNet is particularly significant for portable device design. Its lightweight nature (low parameter count and GFlops) makes it suitable for deployment in resource-constrained environments like bedside ultrasound, where traditional heavy models (e.g., ViT-based or large CNNs) are impractical. The results suggest that HLNet can achieve segmentation performance comparable to fully supervised models trained on full datasets, using only a fraction of the labeled data.