A Composite Model of Wound Segmentation Based on Traditional Methods and Deep Neural Networks

Fangzhao Li, Changjian Wang, Xiaohui Liu, Yuxing Peng, Shiyao Jin

TL;DR

This paper introduces a novel composite model for wound segmentation that combines traditional image processing techniques with deep neural networks (DNNs). By leveraging the strengths of both approaches, the model effectively handles complex backgrounds and limited labeled data, achieving improved wound segmentation accuracy.

What Problem Does This Paper Address?

Accurate wound segmentation is crucial for automated wound area measurement and tissue analysis, which are essential for effective wound treatment and management. Traditional manual methods are time-consuming, labor-intensive, and prone to inaccuracies and hygiene issues. While deep learning methods have shown promise in image segmentation, their application to wound segmentation is hindered by the scarcity of labeled wound image datasets. Existing deep learning approaches often require extensive training data, and traditional methods struggle with complex backgrounds and the need for manual feature engineering. This paper aims to bridge this gap by proposing a hybrid approach that mitigates the limitations of both traditional and deep learning methods for wound segmentation.

Methods

The proposed composite model integrates traditional image processing techniques with deep neural networks in a three-step process:

  1. Preprocessing (Skin with Wound Detection): This initial step focuses on highlighting wound features and augmenting the data. It involves a novel "skin with wound detection" algorithm designed to remove environmental backgrounds and simplify the segmentation task. This algorithm dynamically determines Cr thresholds for skin detection, adapting to variations in skin color and wound characteristics. It also employs a marker-based watershed method for further refinement.
  2. Preliminary Wound Segmentation (Deep Neural Networks): The preprocessed images are then fed into a deep neural network for preliminary segmentation. The authors chose the MobileNet architecture for its ability to adjust network complexity and reduce parameters, which is beneficial given the limited training data. They also incorporated skip links and upsampling layers to preserve spatial information.
  3. Semantic Correction: Finally, the segmentation results from the DNN are refined using traditional image processing methods. This step addresses semantic features of wounds, such as filling holes in predicted wound areas and removing minor noise or misidentified regions. This ensures that the segmented wounds are realistic and adhere to known wound characteristics.

Datasets

The authors collected a dataset of 950 wound images from the internet and hospitals. These images were manually labeled using a semi-interactive wound labeling software. The dataset includes images of various wound types (diabetic foot complications, pressure ulcers, burns) and locations, with varying lighting conditions and backgrounds. The dataset was split into training (760 images) and testing (190 images) sets.

Key Insights and Ideas

  • Hybrid Approach Synergy: Combining traditional image processing (for background removal and semantic correction) with deep learning (for feature extraction and segmentation) offers a powerful synergy, overcoming the individual limitations of each.
  • Dynamic Thresholding for Skin Detection: The proposed dynamic thresholding approach for skin detection is more robust than fixed threshold methods, effectively handling variations in skin color and complex backgrounds.
  • Importance of Preprocessing: Preprocessing the images to remove complex backgrounds significantly simplifies the segmentation task for the DNN and improves accuracy, even with limited training data.
  • Train-Test Group Optimization: The study highlights the critical importance of selecting an optimal train-test group. Using preprocessed images for both training and testing, and augmenting the training set with raw images, yielded the best segmentation results.
  • Parameter Reduction with MobileNet: Utilizing MobileNet with parameter reduction techniques (e.g., interpolation layers) allows for efficient DNNs that can achieve comparable accuracy to larger models with significantly fewer parameters, crucial for datasets with limited training samples.

Key Results

The proposed composite model demonstrated strong performance:

  • The complete model (preprocessing + DNN + postprocessing) achieved the best IoU (Intersection over Union) scores, outperforming models that relied solely on deep learning or only preprocessing.
  • The IoU of the preprocessing model with a depth multiplier (DM) of 0.5 was comparable to the original networks with DM = 1.0, indicating effective parameter reduction without significant loss of precision.
  • The complete model with DM = 1.0 showed a 1.5% improvement in IoU compared to the original networks.
  • The best train-test group configuration (Fourth Group: training with raw and preprocessed data, testing with preprocessed data) yielded the highest segmentation accuracy.
  • Qualitative results showed that the model effectively segmented wounds, even in images with complex backgrounds, and the semantic correction step improved the realism of the segmented areas.

Limitations and Future Work

The authors acknowledge several limitations:

  • Nonskin Background Misjudgment: The skin with wound detection algorithm still occasionally misjudges nonskin backgrounds as skin, which can lead to misclassification of these areas as wounds in the final results. Further improvement of this algorithm is needed.
  • Over-Smoothing of Segmentation: The use of a 16-fold upsampling in the DNN resulted in overly smooth segmentation boundaries compared to the ground truth.
  • Database Dependency: The traditional algorithm's parameter values are determined based on the statistics of the collected database, suggesting a potential dependency on the dataset's characteristics.
  • Image Size Requirement: The current model requires images larger than 512x512 pixels.

Future work suggested includes:

  • Further refining the skin with wound detection algorithm to reduce nonskin background misjudgments.
  • Investigating the use of Conditional Random Fields (CRF) to further refine segmentation boundaries and address the over-smoothing issue.
  • Exploring methods to improve robustness to variations in image statistics and potentially reduce the dependency on specific database characteristics.
  • Expanding the dataset to include more diverse wound types and backgrounds.

Summary updated

Abstract

Wound segmentation plays an important supporting role in the wound observation and wound healing. Current methods of image segmentation include those based on traditional process of image and those based on deep neural networks. The traditional methods use the artificial image features to complete the task without large amounts of labeled data. Meanwhile, the methods based on deep neural networks can extract the image features effectively without the artificial design, but lots of training data are required. Combined with the advantages of them, this paper presents a composite model of wound segmentation. The model uses the skin with wound detection algorithm we designed in the paper to highlight image features. Then, the preprocessed images are segmented by deep neural networks. And semantic corrections are applied to the segmentation results at last. The model shows a good performance in our experiment.
BibTeX
@article{li2018composite,
  author = {Fangzhao Li and Changjian Wang and Xiaohui Liu and Yuxing Peng and Shiyao Jin},
  title = {{A Composite Model of Wound Segmentation Based on Traditional Methods and Deep Neural Networks}},
  journal = {Comput. Intell. Neurosci.},
  year = {2018},
  doi = {10.1155/2018/4149103},
}
Figure