Source-linked AI summary
Fully Automatic Wound Segmentation with Deep Convolutional Neural Networks
Chuanbo Wang, DM Anisuzzaman, Victor Williamson, Mrinal Kanti Dhar, Behrouz Rostami, Jeffrey Niezgoda, Sandeep Gopalakrishnan, Zeyun Yu
TL;DR
Wound segmentation from images addresses healthcare burdens and the need for accurate wound-area measurement. The paper combines MobileNetV2 with connected component labelling, achieving a Dice score increase over VGG16 and supporting mobile application.
Problem
Wounds burden healthcare systems, while image-based segmentation and accurate wound-area measurement support wound-care evaluation and management.
Method
The paper develops a wound-segmentation framework based on MobileNetV2 and connected component labelling, using specialist-annotated wound images.
Results
90.47% Dice score was achieved with the proposed method versus 81.03% for VGG16 on the authors’ dataset.
Takeaways & Limitations
The method is efficient and lightweight, taking less training time and supporting application to mobile devices.
Takeaways & Limitations
Prior wound-segmentation methods are limited by image resolution, manually tuned parameters, and empirically handcrafted features.
Abstract
from arXiv · showhide
Acute and chronic wounds have varying etiologies and are an economic burden to healthcare systems around the world. The advanced wound care market is expected to exceed $22 billion by 2024. Wound care professionals rely heavily on images and image documentation for proper diagnosis and treatment. Unfortunately lack of expertise can lead to improper diagnosis of wound etiology and inaccurate wound management and documentation. Fully automatic segmentation of wound areas in natural images is an important part of the diagnosis and care protocol since it is crucial to measure the area of the wound and provide quantitative parameters in the treatment. Various deep learning models have gained success in image analysis including semantic segmentation. Particularly, MobileNetV2 stands out among others due to its lightweight architecture and uncompromised performance. This manuscript proposes a novel convolutional framework based on MobileNetV2 and connected component labelling to segment wound regions from natural images. We build an annotated wound image dataset consisting of 1,109 foot ulcer images from 889 patients to train and test the deep learning models. We demonstrate the effectiveness and mobility of our method by conducting comprehensive experiments and analyses on various segmentation neural networks.
Introduction
Wound segmentation addresses the clinical and documentation burden of measuring wounds, while prior methods face limitations in robustness, tuning, pathology coverage, and dataset size. This paper proposes a specialist-annotated, fully automatic MobileNetV2-based framework designed to balance segmentation accuracy with computational efficiency.
- Clinical motivation: Wound area measurement supports chronic-wound monitoring and intervention decisions, but manual measurement is time-consuming and often inaccurate.The motivation links quantitative wound measurement to evaluating healing trajectories and determining future interventions.
- Limitations of prior work: Traditional wound-segmentation methods rely on handcrafted features and manually tuned parameters that can be affected by pigmentation, illumination, and image resolution.The introduction also notes that these approaches may not guarantee optimal results and can struggle with severe pathologies and rare cases.
- Limitations of prior work: Existing FCN approaches also showed difficulty distinguishing small wounds or irregular borders, while one study used semi-automatic watershed annotations rather than specialist labels.Smooth contour tendencies limited segmentation accuracy for some wound shapes, and watershed-generated labels represented the algorithm’s labeling behavior.
- This paper: The authors build a large wound-image dataset with segmentation annotations created by wound specialists and describe it as the largest focused wound-segmentation dataset to their knowledge.The paper’s contributions identify specialist annotation and dataset scale as central components of the work.
- This paper: The proposed framework uses MobileNetV2 to provide fully automatic wound segmentation while balancing computational efficiency and accuracy.The network is described as lightweight because it uses significantly fewer parameters during training.
Dataset
The dataset contains clinically collected foot-ulcer images captured under uncontrolled conditions and prepared for fixed-size deep-learning inputs. A localization stage identifies wound regions before image patches are padded and used for segmentation training and testing.
- Image acquisition: The images were captured with a Canon SX 620 HS digital camera and iPad Pro under uncontrolled illumination and varied backgrounds.Figure 1 presents sample images from the dataset.
- Preprocessing: A YOLOv3 localization model was trained on 1,010 images, with augmentation producing 3,645 training images and 405 testing images.The localization model was trained using manually labeled bounding boxes in YOLO format.
- Preprocessing: At an IoU rate of 0.5 and non-maximum suppression of 1.00, the localization model achieved a mean Average Precision of 0.939.The model was trained for 273 epochs with batch size 8.
- Preprocessing: Localized wound patches were cropped from bounding boxes and zero-padded to a fixed size of 224 pixels by 224 pixels.Fixed-size inputs were required by the deep-learning model.
- Annotations: The segmentation masks were manually annotated, then reviewed and verified by wound-care specialists from the collaborating clinic.The dataset initially focused on foot ulcers because their smaller size made pixel-wise annotation easier and less time-consuming.
- Scope: The dataset’s current scope is limited to foot-ulcer images, with future expansion planned for venous leg, pressure, and surgery wounds and non-wound references.The authors identify broader wound-type coverage as a future direction.
Methods
The method uses a MobileNetV2-based encoder-decoder with depth-separable convolutions, transfer learning, and connected-component post-processing for wound segmentation. It evaluates segmentation with Precision, Recall, and Dice and compares multiple neural-network architectures.
- Network architecture: MobileNetV2 is used in an encoder-decoder network to segment wound regions from images.The encoder repeatedly applies depth-separable convolution blocks, while the decoder uses multiscale feature pooling, concatenation, convolutional refinement, and bilinear up-sampling.
- Network architecture: Each depth-separable block combines a 3 × 3 depth-wise convolution with a 1 × 1 point-wise convolution, reducing computational cost relative to conventional convolutions.The layers are followed by batch normalization and Relu6 activation; depth-separable convolutions are described as suitable for resource-limited mobile or embedded applications.
- Training: Transfer learning initializes MobileNetV2 with weights pre-trained on the Pascal VOC segmentation dataset before wound-segmentation training.The authors state that this training strategy helps weights converge faster and better.
- Post-processing: Post-processing thresholds grayscale masks at 127, fills holes, and removes small connected components to produce final segmentation masks.Connected-component labelling addresses holes caused by abnormal tissue and removes small false-positive regions using adaptive thresholds.
- Evaluation: Segmentation performance is evaluated using Precision, Recall, and the Dice coefficient across several neural-network models.The experiments use data augmentation and preprocessing and compare the proposed model with architectures including FCN-VGG-16.
Discussion
The discussion compares wound-segmentation networks and attributes MobileNetV2’s performance and efficiency partly to its deeper architecture and residual skip connections. The method also requires fewer parameters and computational resources than several alternatives.
- Dice score increased from 81.03% with VGG16 to 90.47% with the proposed method on the authors’ dataset.
- MobileNetV2 uses a deeper architecture and residual skip connections that increase its capacity to handle wound-image variation.The discussion identifies variation in tissue, lighting, and perspective as challenges in chronic-wound images.
- U-Net significantly outperformed SegNet in mean Dice score, consistent with the role of skip connections in improving localization.U-Net combines encoder feature maps with decoder outputs to preserve contextual and localization information at higher resolution.
- The adopted MobileNetV2 has only a fraction of the trainable parameters of U-Net, VGG16, and Mask-RCNN.The smaller model trains faster and can run on mobile devices with less memory and limited computational power.
- MobileNetV2 can accept higher-resolution input images with less memory and computational power than the other compared models.
Conclusions
The study addresses automated chronic foot-ulcer segmentation using a dataset built by the authors and evaluates MobileNetV2 with connected component labelling against several neural networks. The proposed model demonstrated effectiveness and mobility, achieved the highest Dice score on the Medetec Wound Dataset, and is intended to be strengthened with more data and a multi-stream architecture.
- The study targets automated segmentation of chronic foot ulcers using deep learning on a dataset built by the authors.
- The experiments compare SegNet, VGG16, U-Net, Mask-RCNN, and the proposed MobileNetV2-based model with connected component labelling.
- The proposed method demonstrated effectiveness and mobility in image segmentation through a fully convolutional architecture with depth-wise separable convolutional layers.
- The model achieved the highest Dice score when tested on foot-ulcer images from the publicly available Medetec Wound Dataset.
- The authors plan to improve robustness and prediction accuracy by adding more data and developing a multi-stream architecture that separately extracts shape features.