Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Region of Interest Focused Triple UNet Architecture for Skin Lesion Segmentation

Guoqing Liu, Yu Guo, Caiying Wu, Guoqing Chen, Barintag Saheya, Qiyu Jin

TL;DR

This work addresses the challenging task of skin lesion segmentation with irregular borders by introducing Triple-UNet, a three-stage UNet-based architecture. It incorporates a region of interest enhancement module (ROIE) that leverages the first network's score map to enrich the input for the second network, followed by a third network that refines the final segmentation via multiplicative fusion, with deep supervision applied at each stage. The encoder uses depthwise separable convolutions and channel attention to reduce parameters, while three decoders—with strategic skip connections—produce progressively accurate score maps. On ISIC-2018, Triple-UNet achieves state-of-the-art Dice and mIoU scores, outperforming competitors like DoubleUNet and UNet variants, and ablation studies confirm the efficacy of ROIE and the three-network arrangement. This approach demonstrates strong potential for precise, ROI-focused medical image segmentation with improved robustness to lesion variability and boundary ambiguity.

Abstract

Skin lesion segmentation is of great significance for skin lesion analysis and subsequent treatment. It is still a challenging task due to the irregular and fuzzy lesion borders, and diversity of skin lesions. In this paper, we propose Triple-UNet to automatically segment skin lesions. It is an organic combination of three UNet architectures with suitable modules. In order to concatenate the first and second sub-networks more effectively, we design a region of interest enhancement module (ROIE). The ROIE enhances the target object region of the image by using the predicted score map of the first UNet. The features learned by the first UNet and the enhanced image help the second UNet obtain a better score map. Finally, the results are fine-tuned by the third UNet. We evaluate our algorithm on a publicly available dataset of skin lesion segmentation. Experiments show that Triple-UNet outperforms the state-of-the-art on skin lesion segmentation.

A Region of Interest Focused Triple UNet Architecture for Skin Lesion Segmentation

TL;DR

This work addresses the challenging task of skin lesion segmentation with irregular borders by introducing Triple-UNet, a three-stage UNet-based architecture. It incorporates a region of interest enhancement module (ROIE) that leverages the first network's score map to enrich the input for the second network, followed by a third network that refines the final segmentation via multiplicative fusion, with deep supervision applied at each stage. The encoder uses depthwise separable convolutions and channel attention to reduce parameters, while three decoders—with strategic skip connections—produce progressively accurate score maps. On ISIC-2018, Triple-UNet achieves state-of-the-art Dice and mIoU scores, outperforming competitors like DoubleUNet and UNet variants, and ablation studies confirm the efficacy of ROIE and the three-network arrangement. This approach demonstrates strong potential for precise, ROI-focused medical image segmentation with improved robustness to lesion variability and boundary ambiguity.

Abstract

Skin lesion segmentation is of great significance for skin lesion analysis and subsequent treatment. It is still a challenging task due to the irregular and fuzzy lesion borders, and diversity of skin lesions. In this paper, we propose Triple-UNet to automatically segment skin lesions. It is an organic combination of three UNet architectures with suitable modules. In order to concatenate the first and second sub-networks more effectively, we design a region of interest enhancement module (ROIE). The ROIE enhances the target object region of the image by using the predicted score map of the first UNet. The features learned by the first UNet and the enhanced image help the second UNet obtain a better score map. Finally, the results are fine-tuned by the third UNet. We evaluate our algorithm on a publicly available dataset of skin lesion segmentation. Experiments show that Triple-UNet outperforms the state-of-the-art on skin lesion segmentation.
Paper Structure (15 sections, 4 equations, 5 figures, 2 tables)

This paper contains 15 sections, 4 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. CNN Models
  4. Application of Attention Mechanism
  5. Skin Lesion Segmentation
  6. Method
  7. Encoder
  8. Decoder
  9. Connectivity structure among sub-networks
  10. Experiment
  11. Datasets
  12. Implementation details
  13. Comparison with State-of-the-Arts
  14. Ablation Study
  15. Conclusion

Figures (5)

  • Figure 1: Block diagram of the proposed Triple-UNet architecture.
  • Figure 2: The architecture of ROIE block.
  • Figure 3: Visual comparison of input images of NETWORK 2 in Triple-UNet and DoubleUNet. (a) is the input image and (d) is its corresponding ground truth. (b) and (c) are the inputs to NETWORK 2 in the corresponding algorithm. (e) and (f) are the prediction mask of DoubleUNet and the prediction mask of Triple-UNet, respectively.
  • Figure 4: Qualitative results on Lesion boundary segmentation dataset from ISIC-2018 2019Skin2018The
  • Figure 5: Segmentation results on ISIC-2019 ISIC2019. BCC is basal cell carcinoma. AK is actinic keratosis. DF is dermatofibroma. VASC is vascular lesion. SCC is squamous cell carcinoma.