Scholarlink

<p>Diabetic foot ulcers (DFUs) are a serious complication associated with diabetes mellitus, as they can develop into infection, hospitalization, and ultimately lower-extremity amputation if they are not identified and monitored soon enough. Therefore, properly segmenting the ulcer regions from clinical images is important for automating the assessment of wounds and planning for treatment options. However, properly segmenting DFUs is a challenge due to their irregular wound boundary morphology, the variability in color and texture of the tissue, and the complex clinical background surrounding the wound. In this research, we describe a novel hybrid Mamba U-Net++ architecture for the segmentation of DFUs.The system that is being proposed leverages two main technologies: Mamba modules and U-Net. The Mamba module excels at modeling contextual cues over long distances while the U-Net model is one of many that utilizes multi-scale feature fusion capabilities. These technologies can be used together in different ways, either via vertical or horizontal skip connections, to create a nested approach to improving feature aggregation. Additionally, a bi-directional Mamba-based spatial block is incorporated into the deeper layers of the network to better model long-range spatial dependencies across the full image. To handle the pronounced class imbalance typically observed in medical segmentation tasks, training is performed using a hybrid loss formulation combining Binary Cross-Entropy (BCE) and Focal Tversky loss. The model is evaluated on the DFUC2022 and FUSC2021 benchmark datasets. On FUSC2021, the proposed method achieves a Dice score of 0.9140 and an Intersection-over-Union (IoU) of 0.8416, while on DFUC2022 it attains a Dice score of 0.8427 and IoU of 0.7281. These results surpass those obtained by standard U-Net and U-Net++ baselines. The findings suggest that incorporating state-space modeling within convolutional segmentation frameworks enhances both prediction accuracy and robustness in automated diabetic wound analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid Mamba and UNet plus plus framework for robust diabetic foot ulcer segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Daniel Willson Velladurai,
  • R. Venkatesan,
  • Saswati Debnath,
  • Jaydeep Kishore,
  • Varun Tiwari

摘要

Diabetic foot ulcers (DFUs) are a serious complication associated with diabetes mellitus, as they can develop into infection, hospitalization, and ultimately lower-extremity amputation if they are not identified and monitored soon enough. Therefore, properly segmenting the ulcer regions from clinical images is important for automating the assessment of wounds and planning for treatment options. However, properly segmenting DFUs is a challenge due to their irregular wound boundary morphology, the variability in color and texture of the tissue, and the complex clinical background surrounding the wound. In this research, we describe a novel hybrid Mamba U-Net++ architecture for the segmentation of DFUs.The system that is being proposed leverages two main technologies: Mamba modules and U-Net. The Mamba module excels at modeling contextual cues over long distances while the U-Net model is one of many that utilizes multi-scale feature fusion capabilities. These technologies can be used together in different ways, either via vertical or horizontal skip connections, to create a nested approach to improving feature aggregation. Additionally, a bi-directional Mamba-based spatial block is incorporated into the deeper layers of the network to better model long-range spatial dependencies across the full image. To handle the pronounced class imbalance typically observed in medical segmentation tasks, training is performed using a hybrid loss formulation combining Binary Cross-Entropy (BCE) and Focal Tversky loss. The model is evaluated on the DFUC2022 and FUSC2021 benchmark datasets. On FUSC2021, the proposed method achieves a Dice score of 0.9140 and an Intersection-over-Union (IoU) of 0.8416, while on DFUC2022 it attains a Dice score of 0.8427 and IoU of 0.7281. These results surpass those obtained by standard U-Net and U-Net++ baselines. The findings suggest that incorporating state-space modeling within convolutional segmentation frameworks enhances both prediction accuracy and robustness in automated diabetic wound analysis.