[Paper Review] UCM-Net: A Lightweight and Efficient Solution for Skin Lesion Segmentation using MLP and CNN

Chunyu Yuan, Dongfang Zhao+1 more|arXiv (Cornell University)|Oct 14, 2023

Cutaneous Melanoma Detection and Management4 citations

TL;DR

UCM-Net proposes a lightweight, efficient deep learning model for skin lesion segmentation by integrating Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN) via a novel UCM-Net Block, achieving state-of-the-art performance with under 50KB parameters and less than 0.05 GLOPs, enabling deployment on mobile health devices.

ABSTRACT

Skin cancer poses a significant public health challenge, necessitating efficient diagnostic tools. We introduce UCM-Net, a novel skin lesion segmentation model combining Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN). This lightweight, efficient architecture, deviating from traditional UNet designs, dramatically reduces computational demands, making it ideal for mobile health applications. Evaluated on PH2, ISIC 2017, and ISIC 2018 datasets, UCM-Net demonstrates robust performance with fewer than 50KB parameters and requires less than 0.05 Giga Operations Per Second (GLOPs). Moreover, its minimal memory requirement is just 1.19MB in CPU environment positions. It is a potential benchmark for efficiency in skin lesion segmentation, suitable for deployment in resource-constrained settings. In order to facilitate accessibility and further research in the field, the UCM-Net source code is https://github.com/chunyuyuan/UCM-Net.

Motivation & Objective

• Address the challenge of deploying high-accuracy skin lesion segmentation models on resource-constrained mobile and point-of-care devices.
• Overcome the limitations of existing deep learning models that require high computational resources and large parameter counts.
• Develop a hybrid architecture combining CNN and MLP to enhance feature learning while minimizing model size and inference cost.
• Improve segmentation accuracy on complex skin lesion images with diverse appearances, poor boundaries, and imaging artifacts.
• Enable real-time, on-device skin cancer diagnosis through a highly efficient, low-footprint model suitable for portable medical devices.

Proposed method

• Propose the UCM-Net Block, a hybrid module that integrates CNN and MLP components to enhance feature representation with reduced parameter count.
• Design a lightweight U-Net-based encoder-decoder architecture using multiple UCM-Net Blocks to maintain high segmentation performance with minimal parameters.
• Introduce a Group Loss function to improve training stability and segmentation accuracy by jointly optimizing multiple loss components.
• Optimize model efficiency by reducing depth and width of the network while maintaining feature hierarchy and spatial resolution.
• Utilize channel-wise and depthwise separable convolutions implicitly through the UCM-Net Block to reduce FLOPs and parameter count.
• Train and fine-tune the model on ISIC2017 and ISIC2018 datasets using standard data augmentation and learning rate scheduling.

Experimental results

Research questions

• RQ1 Can a hybrid CNN-MLP architecture achieve superior skin lesion segmentation performance with drastically reduced model size and computational cost?
• RQ2 How does the proposed UCM-Net Block compare to standard convolutional blocks in terms of parameter efficiency and segmentation accuracy?
• RQ3 To what extent does the Group Loss function improve segmentation performance compared to standard loss functions in skin lesion segmentation?
• RQ4 Can a model with fewer than 50KB parameters and less than 0.05 GLOPs achieve state-of-the-art performance on benchmark skin lesion segmentation datasets?
• RQ5 Is the proposed model suitable for deployment on mobile and edge devices due to its low inference cost and minimal memory footprint?

Key findings

• UCM-Net achieves a mean Intersection over Union (mIoU) of 80.63% and a mean Dice Similarity Coefficient (mDice) of 87.64% on the ISIC2017 dataset when using Group Loss, surpassing the baseline U-Net and other state-of-the-art models.
• The model has only 49,932 parameters and 0.0465 GFLOPs, making it the first skin lesion segmentation model with under 50KB parameters and less than 0.05 GLOPs.
• UCM-Net outperforms EGE-UNet in both mIoU and mDice while using significantly fewer parameters and lower computational cost, even when both models use the same number of channels.
• The ablation study confirms that replacing standard convolutional blocks with the UCM-Net Block improves performance despite a drastic reduction in parameters, with mIoU increasing from 78.48% to 79.76%.
• The inclusion of Group Loss further boosts performance, increasing mIoU from 79.76% to 80.63% and mDice from 86.94% to 87.64%, demonstrating its effectiveness in optimizing segmentation quality.
• Visual results in Figure 5 show that UCM-Net's segmentation outputs closely match ground truth annotations, indicating high precision and boundary adherence on complex lesion cases.