Replication Data for: Improving Objective Wound Assessment: \"Fully-automated wound tissue segmentation using Deep Learning on mobile devices\"

Background: The composition of tissue types present within a wound is a useful indicator of its healing progression and could be helpful in guiding its treatment. Additionally, this measure is clinically used in wound healing tools (e.g. BWAT) to assess risk and recommend treatment. However, the identification of wound tissue and the estimation of their relative composition is highly subjective and variable. This results in incorrect assessments being reported, leading to downstream impacts including inappropriate dressing selection, failure to identify wounds at risk of not healing, or failure to make appropriate referrals to specialists. Objective: To measure inter-and intra-rater variability in manual tissue segmentation and quantification among a cohort of wound care clinicians. To determine if an objective assessment of tissue types (i.e., size, amount) can be achieved using a deep convolutional neural network that predicts wound tissue types. The proposed objective measurement by machine learning model’s performance is reported in terms of mean intersection over union (mIOU) between model prediction and the ground truth labels. Finally, to compare the performance of the model wound tissue identification by a cohort of wound care clinicians. Methods: A dataset of 58 anonymized wound images of various types of chronic wounds from Swift Medical’s Wound Database was used to conduct the inter-rater and intra-rater agreement study. The dataset was split into 3 subsets, with 50% overlap between subsets to measure intra-rater agreement. Four different tissue types (epithelial, granulation, slough and eschar) within the wound bed were independently labelled by the 5 wound clinicians using a browser-based image annotation tool. Each subset was labelled at one-week intervals. Inter-rater and intra rater agreement was computed. Next, two separate deep convolutional neural networks architectures were developed for wound segmentation and tissue segmentation and are used in sequence in the proposed workflow. These models were trained using 465,187 wound image-label pairs and 17,000 image-label pairs respectively. This is by far the largest and most diverse reported dataset of labelled wound images used for training deep learning models for wound and wound tissue segmentation. This allows our models to be robust, unbiased towards skin tones and generalize well to unseen data. The deep learning model architectures were designed to be fast and nimble to allow them to run in near real-time on mobile devices. Results: We observed considerable variability when a cohort of wound clinicians was tasked to label the different tissue types within the wound using a browser-based image annotation tool. We report poor to moderate inter-rater agreement in identifying tissue types in chronic wound images. A very poor Krippendorff alpha value of 0.014 for inter-rater variability when identifying epithelization has been observed, while granulation is most consistently identified by the clinicians. The intra-rater ICC(3,1) (Intra-Class Correlation) however indicates raters are relatively consistent when labelling the same image multiple times over a period of time. Our deep learning models achieved a mean intersection over union (mIOU) of 0.8644 and 0.7192 for wound and tissue segmentation respectively. A cohort of wound clinicians, by consensus, rated 91% of the tissue segmentation results to be between fair and good in terms of tissue identification and segmentation quality. Conclusions: Our inter-rater agreement study validates that clinicians may exhibit considerable variability when identifying and visually estimating tissue proportion within the wound bed. The proposed deep learning model provides objective tissue identification and measurements to assist clinicians in documenting the wound more accurately. Our solution works on off-the-shelf mobile devices and was trained with the largest and most diverse chronic wound dataset ever reported and leading to a robust model when deployed. The proposed solution brings us a step closer to more accurate wound documentation and may lead to improved healing outcomes when deployed at scale.研究背景：伤口内各类组织的构成情况是判断伤口愈合进程的有效指标，也可为伤口治疗方案的制定提供参考。此外，该指标已被应用于临床伤口愈合评估工具（如BWAT）中，用于评估伤口风险并推荐治疗方案。然而，伤口组织的识别及其相对占比的估算具有高度的主观性与变异性，这可能导致评估结果出现偏差，进而引发一系列后续问题，例如敷料选择不当、无法识别存在愈合不良风险的伤口，或是未能及时将患者转诊至专科医师处。 研究目的：首先，评估一组伤口护理临床医师在手工进行伤口组织分割与定量分析时的组间与组内变异程度；其次，探究能否通过预测伤口组织类型的深度卷积神经网络（deep convolutional neural network）实现组织类型（即尺寸、占比）的客观评估，本研究以模型预测结果与真实标注（ground truth labels）之间的平均交并比（mean intersection over union, mIOU）来报告机器学习模型的客观测量性能；最后，对比机器学习模型与伤口护理临床医师在伤口组织识别任务中的表现。 研究方法：本研究从Swift Medical伤口数据库中获取了58张经过匿名化处理的各类慢性伤口图像，用于开展组间与组内一致性研究。将该数据集划分为3个子集，子集间存在50%的重叠，以用于评估组内一致性。5名伤口护理临床医师使用基于浏览器的图像标注工具，独立对伤口床内的4种不同组织类型（上皮组织、肉芽组织、腐肉、焦痂）进行标注。所有子集均以每周为间隔完成标注，并计算了组间与组内一致性系数。随后，本研究分别构建了用于伤口分割与组织分割的两个独立深度卷积神经网络架构，并将二者按顺序集成至所提出的工作流中。这两个模型分别使用465187对伤口图像-标注对与17000对图像-标注对进行训练。据现有文献报道，本研究所使用的标注伤口图像数据集是规模最大、多样性最高的用于训练伤口及伤口组织分割深度学习模型的数据集。该数据集可确保模型具备良好的鲁棒性，不受肤色偏倚影响，并可在未知数据上实现良好的泛化能力。本研究设计的深度学习模型架构兼具快速性与灵活性，可在移动设备上实现近乎实时的推理运行。 研究结果：当伤口护理临床医师使用基于浏览器的图像标注工具对伤口内的不同组织类型进行标注时，我们观察到了显著的变异性。本研究发现，临床医师在慢性伤口图像的组织类型识别任务中，组间一致性处于较差至中等水平。在识别上皮化组织时，组间变异的克里彭多夫阿尔法系数（Krippendorff alpha）仅为0.014，一致性极差；而肉芽组织是临床医师识别一致性最高的组织类型。组内组内相关系数（Intra-Class Correlation, ICC(3,1)）结果则显示，临床医师在一段时间内对同一图像进行多次标注时，其结果具有相对较好的一致性。本研究的深度学习模型在伤口分割与组织分割任务中分别实现了0.8644与0.7192的平均交并比（mIOU）。经全体伤口护理临床医师共识评定，91%的组织分割结果在组织识别与分割质量方面被评为“尚可”至“良好”等级。 研究结论：本研究的组间一致性研究证实，临床医师在识别并视觉估算伤口床内组织占比时，可能存在显著的变异性。本研究提出的深度学习模型可实现组织类型的客观识别与定量测量，辅助临床医师更精准地完成伤口文档记录工作。本解决方案可在通用商用移动设备上运行，且基于目前已报道的规模最大、多样性最高的慢性伤口数据集进行训练，部署后可获得鲁棒性极佳的模型性能。该研究成果使我们在实现精准伤口文档记录的道路上迈进了一步，若大规模部署，有望改善伤口愈合结局。