💻 computer science

DualResolution Residual Architecture with Artifact Suppression for Melanocytic Lesion Segmentation

This paper proposes a novel dual-resolution residual architecture enhanced with artifact suppression and multi-task training strategies to achieve precise, pixel-accurate segmentation of melanocytic lesions in dermoscopic images, outperforming traditional baselines in boundary localization and clinical relevance.

Original authors: Vikram Singh, Kabir Malhotra, Rohan Desai, Ananya Shankaracharya, Priyadarshini Chatterjee, Krishnan Menon Iyer

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vikram Singh, Kabir Malhotra, Rohan Desai, Ananya Shankaracharya, Priyadarshini Chatterjee, Krishnan Menon Iyer

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 trace the outline of a very tricky shape drawn on a piece of paper. But there are three problems:

The shape has fuzzy, wiggly edges that look a lot like the background.
There are hairs, ink marks, and shiny bubbles stuck to the paper that look like part of the shape.
You only have a few examples of this shape to learn from.

This is exactly the challenge doctors face when looking at dermoscopic images (close-up photos of skin moles) to find melanoma (skin cancer). They need to draw a perfect line around the mole to measure it and decide if it's dangerous. If the line is even slightly off, the measurements change, and the diagnosis could be wrong.

The paper introduces a new computer program called "Our method" (a dual-resolution ResNet-inspired architecture) designed specifically to solve this "tricky tracing" problem. Here is how it works, using simple analogies:

1. The Two-Stream Team: The "Detail Detective" and the "Big Picture Thinker"

Most old computer programs try to do everything at once, which often makes them miss small details or get confused by the big picture. This new method uses a two-person team that works together constantly:

The Detail Detective (Full-Resolution Stream): This team member never zooms out. They keep looking at the image at its highest possible resolution. Their job is to spot the tiny, jagged edges of the mole and the fine hairs. They ensure the final outline is pixel-perfect.
The Big Picture Thinker (Pooled Stream): This team member zooms out to see the whole context. They look at the general shape, the colors, and the texture to figure out, "Is this actually a mole, or just a weird shadow?" They help the Detective understand the bigger context so they don't get tricked by confusing patterns.

How they talk: They don't work in isolation. They pass notes back and forth. The Big Picture Thinker sends clues about the general shape to the Detective, and the Detective sends back sharp, high-quality edge information to the Thinker. This ensures the final outline is both accurate and contextually correct.

2. The "Noise Canceling" Headphones (Artifact Suppression)

Dermoscopic photos are often messy. They have stray hairs, ruler markings, ink dots, and shiny reflections (specular highlights) that can trick a computer into thinking those are part of the mole.

The new system has a special "Artifact Suppression Block." Think of this like noise-canceling headphones. Before the computer tries to trace the mole, this block scans the image and says, "Hey, that's just a hair, ignore it," or "That's a shiny bubble, don't count that." It filters out the distractions so the computer focuses only on the actual skin lesion.

3. The "Smart Lens" (Channel Attention)

Skin comes in many colors, and moles can look very different on light skin versus dark skin. A standard camera might struggle to see the contrast on darker skin tones.

This system uses a Channel Attention Module, which acts like a smart lens that automatically adjusts its focus. If the image is dark, the lens tightens its focus on the specific color contrasts that matter for that skin tone. If the texture is rough, it focuses on the texture. It adapts to the specific "look" of the image to make sure it doesn't miss anything.

4. The "Strict Teacher" (Training with Multiple Goals)

To teach this computer system, the authors didn't just use one rule. They used a multi-task training objective, which is like a strict teacher who grades the student on three different things at once:

The Overlap Score: "Did you cover the right amount of area?" (Dice/Tversky loss).
The Edge Score: "Is your line exactly on the border?" (Boundary loss).
The Consistency Score: "If I rotate the picture or change the lighting slightly, do you still see the same mole?" (Contrastive regularizer).

By punishing the computer if it fails on any of these three, the system learns to be incredibly robust, even when there aren't many training images available.

The Results: Why It Matters

The paper tested this new "two-stream team" against older, standard computer programs (called FCN baselines). The results showed that the new method is significantly better at:

Drawing the line: It creates much sharper, more accurate boundaries.
Handling trouble spots: It works much better on dark skin tones, on body parts with lots of hair (like the head or neck), and on older skin with wrinkles.
Reducing mistakes: It makes fewer errors where it thinks a hair is a mole (false positives) or misses a part of the mole (false negatives).

In summary: This paper presents a specialized computer vision tool that acts like a highly skilled, two-person team equipped with noise-canceling headphones and smart lenses. It is designed specifically to trace the messy, fuzzy outlines of skin moles in photos, filtering out distractions and adapting to different skin types to provide a precise measurement that doctors can trust.

Technical Summary: Dual-Resolution Residual Architecture with Artifact Suppression for Melanocytic Lesion Segmentation

Problem Statement
Accurate segmentation of melanocytic tumors in dermoscopic images is a prerequisite for automated skin cancer screening and clinical decision support. Unlike natural scene segmentation, this task faces unique challenges: low contrast between lesions and surrounding skin, significant texture and color heterogeneity within lesions, and frequent imaging artifacts such as hairs, ink markings, ruler scales, and specular highlights. Furthermore, clinical datasets are often limited in size, and small boundary errors can drastically alter diagnostic measurements and risk stratification. Standard encoder-decoder architectures, while effective for general recognition, often degrade spatial precision due to successive downsampling, making them suboptimal for the pixel-accurate boundary localization required in dermatology.

Methodology
The authors propose "Our method," a novel ResNet-inspired dual-resolution architecture designed to balance robust semantic recognition with precise boundary delineation. The framework consists of the following core components:

Dual-Resolution Streams:
- Full-Resolution Stream: Maintains high-frequency spatial information throughout the encoder to preserve fine-grained boundary details.
- Pooled Stream: Progressively reduces spatial resolution to aggregate multi-scale contextual cues, aiding in the disambiguation of lesion tissue from confounding skin patterns and artifacts.
Full-Resolution Residual Units (FRRUs):
These units tightly couple the two streams. They downsample the full-resolution features, concatenate them with pooled features, and process them through a compact multi-scale convolutional block. The resulting features are upsampled and injected back into the full-resolution stream as residuals. This design allows the pooled stream to evolve robust semantic descriptors while continuously updating the high-resolution stream with contextual information.
Specialized Modules for Dermoscopy:
- Artifact Suppression Branch: A lightweight branch predicts an artifact confidence map (detecting hairs, ink, and highlights) to modulate the residual stream, reducing the influence of corrupted pixels while preserving valid lesion signals.
- Channel-Spatial Attention: A module adapts feature sensitivity to the specific color and texture cues typical of dermoscopic lesions, dynamically reweighting channels and spatial regions.
- Boundary-Aware Residual Injection: An edge head extracts gradient responses from pooled representations, which are projected and upsampled to enforce edge fidelity in the full-resolution stream.
Training Objectives:
To address class imbalance and limited data, the model employs a multi-task loss function:
- Dice–Tversky Loss: Handles class imbalance and controls false positives/negatives.
- Boundary Loss: Implemented as an L1 loss on the signed distance transform (SDT) or banded cross-entropy, explicitly penalizing contour mismatches.
- Contrastive Regularizer: Encourages consistent lesion representations across augmented views to stabilize feature learning.

Key Contributions

Architecture: Introduction of a dual-resolution framework that avoids the spatial precision loss typical of standard downsampling encoders by maintaining a continuous full-resolution stream coupled via FRRUs.
Artifact Handling: Integration of a dedicated artifact suppression block that specifically targets common dermoscopic distractors (hairs, rulers) without requiring complex pre-processing.
Training Strategy: A composite loss function combining region overlap, explicit boundary fidelity, and contrastive regularization, tailored for small, imbalanced clinical datasets.
Performance: The method achieves pixel-accurate segmentation without relying on heavy post-processing or complex pre-training protocols.

Experimental Results
The method was evaluated on public dermoscopic benchmarks and compared against a Fully Convolutional Network (FCN) baseline. Key findings include:

Overall Performance: "Our method" achieved a Dice Coefficient (DC) of 0.88 and Intersection over Union (IoU) of 0.80, significantly outperforming the FCN baseline (DC 0.75, IoU 0.65). The improvement was statistically significant ( $p < 0.001$ ).
Subgroup Analysis: The model demonstrated consistent improvements across skin tones, genders, age groups, and anatomical regions. Notably, it showed substantial gains on dark skin (DC improvement of 0.17) and reduced the performance gap between light and dark skin tones compared to the baseline.
Boundary Fidelity: The method reduced the Average Symmetric Surface Distance (ASSD) by 25–35% and improved the Boundary F1 score by 12–18 percentage points, confirming the efficacy of the boundary-aware components.
Ablation Studies: Removing the artifact suppression module caused a 0.04 drop in IoU on hair-rich regions. Removing boundary injection or loss increased ASSD by 20–30%, validating the necessity of these components for contour accuracy.
Efficiency: While inference time increased slightly (22 ms vs. 12 ms on an RTX 3090) due to the dual-stream processing, the authors note this remains acceptable for batch and near-real-time workflows.

Significance and Claims
The paper claims that "Our method" provides a practical building block for automated melanoma assessment systems. By explicitly addressing the trade-off between semantic context and spatial precision, and by mitigating the impact of common imaging artifacts, the architecture delivers more accurate and reliable lesion masks than standard encoder-decoder baselines. The authors emphasize that the approach is particularly beneficial in challenging scenarios involving low lesion contrast, dense hair, or complex illumination. The results suggest the method supports downstream tasks such as feature extraction, quantitative size measurement, and risk stratification, offering improved fairness across different skin tones and patient demographics. The authors conclude that the architecture achieves a strong balance between precise localization and robust recognition without requiring restrictive pre-training constraints.