💻 computer science

NeviUNet: An Efficient Attention-Enhanced Network for Skin Lesion Segmentation in Total Body Images

NeviUNet is an efficient, attention-enhanced U-Net variant that achieves accurate and low-latency skin lesion segmentation in total body images by integrating Mish activation, Attention Gates, and SE blocks while omitting computationally expensive components.

Original authors: Karolina Dzwonkowska, Piotr Marcin Szczypiński

Published 2026-09-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Karolina Dzwonkowska, Piotr Marcin Szczypiński

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 human skin is a vast, complex landscape, serving as both a protective shield and a mirror to our internal health. Within this landscape, pigmented spots known as moles are common, but some can change in ways that signal a serious threat like melanoma. Because these changes can be subtle and happen over years, doctors rely on long-term monitoring, often using high-resolution cameras to photograph the entire body from every angle. This creates a massive amount of visual data, but for a computer to help a doctor, it must first learn to find and outline these specific spots against the background of normal skin. This process, called segmentation, is difficult because the spots can look very different from one another, sometimes blend into the skin, and are often surrounded by distracting features like hair or uneven lighting.

For years, scientists have tried to teach computers to do this work using deep learning, a type of artificial intelligence that mimics the way the human brain processes images. The most popular tool for this job has been a specific computer architecture known as U-Net, which acts like a sophisticated map-maker, zooming in to see fine details and zooming out to understand the whole picture. However, a persistent problem has emerged: the most accurate computer models are often so heavy and complex that they run too slowly to be useful in a real clinic, while the faster models are often too simple to find the tricky spots accurately. Researchers at the Lodz University of Technology in Poland set out to solve this specific dilemma by creating a new system that is both sharp enough to see the details and light enough to run quickly on standard medical equipment.

The team developed a new computer model they named NeviUNet, designed specifically to analyze images taken by a prototype full-body scanner called the Dermo Studio. This scanner captures forty high-resolution images of a patient's body as they rotate, creating a complete digital record of their skin. To teach their new model, the researchers gathered a collection of these images and had a skin specialist manually draw precise outlines around the moles in each picture. Because the images are so large and detailed, the computer cannot process them all at once, so the team broke them down into smaller, manageable squares. They then trained their AI to look at these squares and decide which pixels belonged to a mole and which belonged to normal skin.

The researchers faced a significant hurdle: in these full-body photos, the actual moles cover only a tiny fraction of the image, while the rest is just skin. This imbalance makes it very easy for a computer to prioritize the majority class and guess that everything is just background skin, effectively ignoring the moles entirely. To prevent this, the team designed their model with a special focus on the rare, important parts of the image. They built the system using a streamlined version of the U-Net structure but added specific, targeted upgrades. They included a mechanism that helps the computer pay attention to the most relevant features and a special mathematical approach that penalizes the computer heavily if it misses a mole, forcing it to be more careful. Crucially, they deliberately left out some of the heavy, complex components found in other advanced models that are known to slow things down without necessarily adding enough value for this specific task.

When they tested their creation, the results were clear. The new NeviUNet model successfully identified and outlined the moles with a high degree of accuracy, significantly outperforming both the basic, older models and the much heavier, more complex models that had been tried before. While the heavier models struggled to run smoothly and often produced inconsistent results, the new model was able to process the images quickly and reliably. In fact, the system was fast enough to analyze the thousands of image squares generated by a single body scan in a timeframe that matches the speed of the scanner itself. This means that in a real-world clinic, a doctor could potentially see the computer's analysis of a patient's entire body almost immediately after the scan is finished, without waiting for slow processing times.

The study also revealed something important about how these computer models should be built. The researchers found that simply making a model bigger or adding more complex layers does not automatically make it better. In their tests, the most complex models available actually performed worse than their new, lighter design, likely because they were too complicated to learn effectively from the limited amount of training data they had. The success of NeviUNet suggests that the best path forward is not to build the biggest possible computer brain, but to build a smarter, more efficient one that is carefully tuned to the specific job it needs to do. By balancing accuracy with speed, this new approach offers a practical way to bring advanced artificial intelligence into the daily practice of skin cancer detection, making it possible to monitor patients more effectively and catch dangerous changes earlier.

Technical Summary: NeviUNet for Skin Lesion Segmentation in Total Body Images

Problem Statement
The paper addresses the challenge of performing efficient and accurate automated segmentation of skin lesions within total body photography acquired under clinical conditions. While deep learning, particularly U-Net architectures, has advanced skin lesion segmentation, a critical trade-off exists between predictive accuracy and computational efficiency. Highly complex models like ResUNet++ often achieve strong performance but incur high computational costs, rendering them unsuitable for resource-constrained or real-time clinical environments. Furthermore, total body images present unique difficulties, including severe class imbalance (where lesions occupy a small fraction of the image area), low contrast, irregular boundaries, and artifacts such as hair. Existing solutions often fail to balance the need for high-resolution analysis with the strict latency requirements of clinical deployment, where segmentation must occur rapidly during or immediately after a scan.

Methodology
The study proposes NeviUNet, a modified U-Net architecture designed to bridge the gap between the lightweight baseline U-Net and the computationally heavy ResUNet++.

Dataset and Preprocessing: The model was trained on a dataset derived from a prototype "Dermo Studio" total-body scanner, which captures high-resolution (20 MP) images of the entire skin surface from multiple angles. Due to the labor-intensive nature of manual annotation by dermatology experts, the dataset was limited to 29 high-resolution images from five patients. To manage GPU memory constraints, images were decomposed into 512×512 pixel patches with 12.5% overlap. A significant preprocessing step involved removing 85% of patches containing only background to mitigate severe class imbalance. The data was split such that one patient's data served as an independent test set, while the remaining data was divided into training (80%) and validation (20%) sets.
Architecture Design: NeviUNet retains the encoder-decoder structure of U-Net but incorporates specific enhancements inspired by ResUNet++ while omitting computationally expensive components:
- Encoder: Consists of four levels with filters increasing from 16 to 128. It utilizes two 3×3 convolutional layers per level followed by Batch Normalization and the Mish activation function (chosen for improved gradient propagation over ReLU).
- Attention Mechanisms: The architecture integrates Squeeze-and-Excitation (SE) blocks after each encoder stage and the bottleneck layer to perform channel-wise attention. Additionally, Attention Gates are applied prior to skip connections to filter encoder features and suppress background noise.
- Omissions: To maintain efficiency, the model explicitly omits residual connections and the Atrous Spatial Pyramid Pooling (ASPP) module, both of which are present in ResUNet++ but contribute significantly to computational load.
- Decoder: Uses transposed convolutions for upsampling (consistent with U-Net) rather than the bilinear upsampling found in ResUNet++.
Loss Function: To address class imbalance and improve boundary delineation, a composite loss function was employed, combining Focal Tversky Loss and Binary Cross Entropy (BCE). The Focal Tversky component heavily penalizes false negatives (using parameters $\alpha=0.8, \beta=0.2$ ) and focuses optimization on low-confidence pixels via a focal parameter $\gamma$ .
Training: The model was trained using the AdamW optimizer with early stopping based on validation Dice-Sørensen scores.

Key Results
The proposed approach was evaluated against a baseline U-Net and ResUNet++ on an independent test set using the Dice-Sørensen Coefficient (DSC) as the primary metric.

Performance: NeviUNet achieved a mean DSC of 0.729, significantly outperforming the best ResUNet++ configuration (0.531) and the baseline U-Net (which failed to detect lesions, yielding a DSC near zero). The proposed model also demonstrated reduced variance in performance compared to ResUNet++.
Computational Efficiency: NeviUNet maintained a lightweight structure with approximately 1.98 million trainable parameters, comparable to the baseline U-Net (1.94 million) and significantly lower than ResUNet++ (4.06 million).
Inference Latency: NeviUNet achieved an inference latency of approximately 32 ms for a batch size of 8 with high stability. In contrast, ResUNet++ exhibited latency exceeding 200 ms with high variance under the same conditions. The baseline U-Net was faster but failed in accuracy.
Training Speed: NeviUNet completed training in 24 minutes (78 epochs), whereas ResUNet++ required 83 minutes (51 epochs).

Significance and Claims
The paper claims that NeviUNet successfully demonstrates that high segmentation accuracy in challenging total body imaging scenarios does not require increasing model complexity or parameter counts. By selectively integrating attention mechanisms (SE blocks, Attention Gates) and the Mish activation function while removing computationally intensive residual and ASPP modules, the authors achieved a model that is both accurate and efficient.

The significance of this work lies in its suitability for real-time clinical deployment. The inference speed (approximately 32 ms for a batch of 8 patches) and low hardware resource requirements make NeviUNet viable for integration into automated total body skin imaging systems like the Dermo Studio scanner. The authors emphasize that their approach offers a practical solution for longitudinal monitoring of melanocytic nevi, enabling reliable lesion detection and tracking without the prohibitive computational costs associated with state-of-the-art heavy architectures. The study concludes that architectural adaptation to specific application constraints is more effective than simply scaling up model capacity.

More like this