💻 computer science

Artifact-Aware Lightweight Dermoscopic Lesion Segmentation with Reproducible Corruption Robustness Evaluation

This paper proposes an artifact-aware lightweight U-shaped segmentation framework and a reproducible corruption benchmark to enhance the robustness and accuracy of dermoscopic lesion segmentation in the presence of acquisition artifacts like hair and specular highlights.

Original authors: Haixiao Gong, Cuijiao Yin, Jie He, Hui Guo, Mei Tang, Linting Meng, Chengchen Gong

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Haixiao Gong, Cuijiao Yin, Jie He, Hui Guo, Mei Tang, Linting Meng, Chengchen Gong

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

The Skin Detective's Dilemma

Imagine you are a detective trying to solve a mystery, but your only clues are blurry, smudged photographs. This is the daily reality for computers trying to spot skin cancer. In the world of medical imaging, there is a special tool called a dermoscope. Think of it as a super-powered magnifying glass that doctors use to look at moles and spots on the skin, peering beneath the surface to see structures invisible to the naked eye. To help doctors, scientists have built "AI detectives"—computer programs designed to automatically trace the outline of these suspicious spots. This process is called "segmentation," and it's like drawing a perfect circle around a suspect in a lineup.

However, real life is messy. When a doctor takes a picture of a mole, the photo often comes with "noise" or "artifacts." These aren't digital glitches; they are real physical things that get in the way. Imagine trying to draw a circle around a puddle while someone is blowing bubbles on the surface, shining a flashlight directly into your eyes, or dropping a few strands of hair across the water. In dermoscopy, these "artifacts" are things like stray hairs, shiny reflections from wet skin, air bubbles trapped under the lens, or the natural lines and wrinkles of the skin itself. These distractions can trick the AI detective, making it think a hair is part of the mole or that a shiny spot is a dangerous edge. The big question scientists are asking is: How do we build a lightweight AI detective that is smart enough to ignore these distractions and still draw the perfect circle, even when the photo is messy?

The Paper's Solution: A Detective with "X-Ray" Vision

This paper introduces a new, clever AI detective called ARL-UNet. The researchers noticed that most current AI models are like students who study perfectly for a test but panic when the teacher adds a few extra scribbles to the page. If the photo has hair or bubbles, these models often get confused and draw the wrong outline. The authors propose a solution that doesn't just try to "clean" the photo before looking at it (which can accidentally remove important clues) but instead teaches the AI to see the distractions and ignore them while it works.

The Three-Part Superpower
The new framework works like a three-step process to handle the mess:

The "Sniffer" (Artifact-Aware Stem): Before the AI even starts looking deep into the image, it takes a quick sniff to find the trouble spots. It uses a mix of simple math tricks (like looking for very bright spots or very dark lines) and a tiny bit of learning to create a "map" of where the hair, bubbles, or glare are. Think of this as the detective putting on special glasses that highlight the bubbles and hairs in red, so they don't get mixed up with the actual crime scene.
The "Efficient Brain" (Robust Encoder): The AI needs to be fast and small enough to run on a mobile phone or a handheld device in a doctor's office. To do this, the authors built a special brain block that looks at the image in different sizes at once. It's like having a team of inspectors: one looks at the big picture, while others zoom in on tiny details. This team works together efficiently to figure out what the mole looks like, even if some parts are covered by a bubble.
The "Gatekeeper" (Mask-Guided Skip Fusion): This is the most important trick. In many AI models, the "shallow" details (the early, blurry layers of the image) are passed directly to the final decision-maker. But if those early layers are full of hair or glare, the final decision gets messed up. The new "Gatekeeper" module acts like a bouncer. It looks at the "Sniffer's" map and says, "Hey, this part of the image has a hair on it; let's not pass that specific detail to the final decision." It filters out the noise while keeping the important edges of the mole.

The "Training Gym" for Messy Photos
To prove their detective was tough, the researchers didn't just test it on perfect photos. They built a new "gym" called a reproducible corrupted benchmark. They took thousands of real medical images and artificially added four types of mess:

Terminal Hair: Strands of hair crossing the image.
Specular Highlights: Bright, shiny reflections.
Air Bubbles: Semi-transparent bubbles that hide parts of the skin.
Skin Lines: The natural texture and wrinkles of the skin.

They created two levels of mess: "Mild" (Severity 1) and "Strong" (Severity 2). Crucially, they kept the "correct answer" (the true outline of the mole) exactly the same, even after adding the mess. This allowed them to measure exactly how much the AI's performance dropped when the photo got dirty.

What They Found
The experiments showed that the new ARL-UNet is a champion in the gym.

Clean Performance: Even on perfect, clean photos, the new AI was just as good as, or slightly better than, the best existing models. It didn't lose its skills just because it was built to handle mess.
Messy Performance: When the photos were covered in hair, bubbles, or glare, the new AI held its ground much better than the others. For example, when tested on images with severe hair or skin lines, it kept its accuracy much higher than older models, which tended to get confused and draw the wrong shapes.
The Training Secret: The researchers tested different ways to train the AI. They found that a "curriculum" approach worked best. This is like teaching a student to walk on a flat floor first, then on a slightly bumpy path, and finally on a very rocky road. If you throw the student onto the rocky road immediately (direct augmentation), they stumble. But if you ease them into it (curriculum training), the AI learns to ignore the mess without forgetting how to see the mole clearly.

The Verdict
The paper suggests that by explicitly teaching the AI to recognize and filter out these specific distractions during the process, rather than trying to remove them beforehand, we can build skin cancer detectors that are both lightweight (fast and small) and incredibly robust. The authors emphasize that while their "messy photo" test is a controlled simulation and not a replacement for real-world clinical trials, it provides a reliable way to measure how well these models will handle the inevitable imperfections of real-life medical imaging. The code for this new detective is available for others to try, offering a promising step toward mobile, point-of-care skin cancer analysis that doesn't break down when the lighting is bad or a hair falls on the lens.

Problem Statement

Automated dermoscopic lesion segmentation is critical for computer-aided skin cancer analysis. While lightweight models are essential for mobile and point-of-care deployment, they often exhibit instability when images contain acquisition artifacts such as terminal hair, specular highlights, skin lines, and air bubbles. These artifacts introduce high-frequency structures, local saturation, and partial occlusions that can be confused with lesion boundaries. Existing lightweight segmentation networks typically optimize for efficiency and clean-image accuracy but fail to explicitly model artifact-prone regions. Furthermore, standard evaluations rely on clean validation images, which do not reflect reliability under real-world acquisition degradations. There is a lack of a fixed, reproducible benchmark to compare robustness against clinically plausible corruptions.

Methodology

The authors propose ARL-UNet, an artifact-aware lightweight U-shaped segmentation framework, alongside a reproducible corrupted benchmark for robustness evaluation.

1. ARL-UNet Architecture

The network is designed to estimate artifact cues before downsampling and use them to modulate skip connections conservatively, rather than removing image regions.

Artifact-Aware Stem (AAS): Located at the input stage before spatial downsampling, the AAS estimates artifact-prone regions using a multi-cue approach. It combines:
- Hand-crafted cues: Laplacian response (high-frequency edges), Bright-region cue (saturated regions), and Dark-region cue (hair-like structures).
- Learned cue: A lightweight branch using depthwise separable convolutions (DSConv) to capture complex artifact appearances.
- These cues are fused into an artifact map ( $A_0$ ) which is used for residual soft suppression of the stem feature, allowing the model to gradually activate suppression during training.
Robust Encoder Block (REB): To maintain a lightweight parameter budget while capturing multi-scale context, the encoder uses depthwise separable convolutions with complementary receptive fields (3×3 and 5×5 or dilated 3×3). It incorporates a channel gating mechanism (squeeze-and-excitation style) to adaptively emphasize informative channels.
Mask-guided Robust Skip Fusion (MRSF): This module addresses the issue of skip connections propagating contaminated shallow features to the decoder. It fuses encoder skip features with decoder features using:
- Semantic Guidance: A coarse lesion mask predicted by an auxiliary head.
- Artifact Guidance: The resized artifact map from the AAS.
- Mechanism: It applies a conservative soft suppression factor to attenuate responses in artifact-prone regions while amplifying them in lesion-consistent regions. The suppression strength is learnable and bounded to prevent over-suppression of true lesion boundaries.
Deep Supervision: Auxiliary prediction heads are attached to multiple decoder stages to provide gradient supervision and serve as semantic guidance for the MRSF module.

2. Reproducible Corrupted Benchmark

To evaluate robustness beyond clean images, the authors constructed a benchmark from fixed ISIC 2017 and ISIC 2018 validation splits.

Design: The benchmark generates corrupted images offline using deterministic seeds, ensuring reproducibility. Ground-truth masks remain unchanged.
Artifacts: Four clinically relevant categories are synthesized at two severity levels (mild-to-moderate and stronger but plausible):
1. Terminal Hair: Dark, thin, curvilinear structures.
2. Specular Highlight: Bright, saturated reflection regions.
3. Air Bubble: Semi-transparent circular/elliptical occlusions.
4. Skin Line: Fine, low-contrast linear patterns.
Metadata: The benchmark records artifact type, severity, generation parameters, and overlap ratios to support controlled analysis.

3. Training Strategies

The study compares three training regimes:

T0 (Clean-only): Standard training without synthetic artifacts.
T1 (Direct Augmentation): Random injection of synthetic artifacts throughout training.
T2 (Curriculum Learning): A staged approach starting with clean-only training, then progressively introducing artifacts (starting with Severity 1, then Severity 2).

Key Contributions

Framework: A novel artifact-aware lightweight segmentation framework (ARL-UNet) that integrates early artifact cue estimation, efficient robust encoding, and mask-guided skip fusion to reduce the propagation of artifact-contaminated features.
Benchmark: A reproducible corrupted dermoscopic benchmark with four clinically relevant artifact categories and two severity levels, providing a standardized protocol for evaluating robustness beyond conventional clean testing.
Systematic Analysis: A comprehensive study of artifact-aware training strategies (clean-only, direct augmentation, and curriculum learning) and their impact on clean-domain accuracy, corrupted-domain robustness, and zero-shot transfer.

Experimental Results

Experiments were conducted on ISIC 2017, ISIC 2018, and the PH2 dataset (for zero-shot evaluation).

Clean-Domain Performance: ARL-UNet achieved competitive state-of-the-art results on clean ISIC 2017 and 2018 benchmarks, outperforming baselines like EMA-UNet in mIoU, Dice, and HD95 while maintaining a lightweight parameter count (~0.147M parameters).
Robustness on Corrupted Benchmarks:
- Under T0 (Clean-only training), ARL-UNet demonstrated superior inherent robustness compared to baselines, indicating the architecture itself mitigates artifact sensitivity.
- Under T2 (Curriculum Learning), the model achieved the highest robustness, particularly under Severity 2 corruptions. It showed significant gains in mIoU over baselines (e.g., +1.29 points over EMA-UNet on ISIC 2017 Severity 2).
- The model performed best on hair and skin-line corruptions, which are challenging due to their boundary-like structures.
Zero-Shot Transfer: Models trained on ISIC datasets and tested directly on PH2 without fine-tuning showed that ARL-UNet (especially with curriculum training) generalizes better than baselines, achieving higher mIoU and lower HD95.
Ablation Studies:
- Removing the AAS or MRSF modules resulted in significant drops in corrupted-domain performance, confirming their critical role in handling artifacts.
- Curriculum training was found to be superior to direct augmentation, as it preserved clean-domain accuracy while improving robustness to severe corruptions.

Significance and Claims

The paper claims to provide a lightweight segmentation framework and a reusable evaluation protocol specifically designed for studying artifact robustness in dermoscopic image segmentation.

Reproducibility: The work emphasizes the creation of a fixed, reproducible benchmark that allows for controlled comparison of artifact-specific degradation, addressing a gap in current literature where robustness is often evaluated on uncontrolled or mixed datasets.
Architectural Insight: The study demonstrates that explicitly modeling artifact-prone regions and using conservative skip fusion is more effective than simply removing artifacts or relying on standard data augmentation.
Practicality: The framework is designed for resource-constrained environments (mobile/point-of-care) while maintaining high accuracy.
Limitations and Future Work: The authors modestly note that the benchmark relies on synthetic artifacts and should not replace real-world clinical validation. They acknowledge that naturally acquired artifacts may be more complex and mixed. Future work is proposed to extend the benchmark with real artifact annotations and multi-center datasets to assess practical generalization.

In summary, the paper argues that for lightweight dermoscopic segmentation to be reliable in clinical settings, models must be evaluated under reproducible, artifact-specific conditions, and that architectural mechanisms to handle these artifacts are as important as the segmentation task itself.