深度學習用於傷口診斷：從急性燙傷到慢性傷口

Abstract

機器學習用在非常多的醫學領域，例如：腫瘤偵測、基因分析、藥物的開發，然而運用於傷口的診斷卻非常的少見。主要的是因為傷口照片取得的限制，傷口的照片不同其他醫學影像是標準化，也無法完全去識別化，拍攝時更無法準確控制病患姿勢，以及控制拍攝的條件(光線及角度)，因此傳統的機器學習，以非常有限的訓練集，儘管在文獻中結果不錯，實際使用卻不如預期。 與傳統的機器學習相比，深度學習可以輸入所有看似不利結果的資訊，並且讓模型在各種環境下皆可以準確診斷。例如：光線的通道，常在傳統的機械學習中被排除，以便使用少量的資料訓練就得到很好的結果，然而以此訓練出的模型，卻只要有稍微光線不良的情況，預測的結果就大打折扣，針對深度學習模型的訓練，則可以把這些因素都一起加入訓練，訓練的情境越多元化，模型實際應用的判斷能力越好，然而深度模型的建立，起初需要大量且系統化標註正確的資料，而正確的標註傷口是十分重要卻又困難。 不同種類的傷口，有不同的診斷需要。例如：急性燙傷的評估，最重要的是燙傷面積佔全身多少比例的體表面積(% TBSA)以及可能需要植皮及清創的深度燙傷；對於慢性傷口，不同組織的組成比例很重要，傷口的絕對面積大小也很重要。上述的這些參數，大致上是兩個任務的衍伸，傷口分割及組織分類。傷口分割是便是那些範圍是傷口，那些是正常皮膚，進一步的可以算出傷口的絕對或相對面積。而組織辨識是為了區分傷口床內以及傷口周圍的不同組織，以得知傷口癒合狀態，肉芽是慢性傷口的一種組織，而深度燙傷也可視為燙傷大範圍中的一種組織。因此這樣概念可運用於急性燙傷也可用於慢性潰瘍。 針對傷口的邊界及周圍組織，使用邊界的標記方式，對於傷口內部的組織，使用以超像素切割後的區域標定方式法。此標記的方式可以用於急性的燙傷及慢性傷口標定。不同深度模型以此系統性標定方法的資料庫訓練，可在傷口分割及組織辨識都有不錯的結果。

Machine learning (ML) has been applied in many medical fields, such as tumor detection, genetic decoding, and drug development. The clinical application for wound assessment is relatively scarce. The reason is mainly because of the limitation of images of wounds. The images of wounds are not standardized and de-identification. Images are usually taken in poor conditions (light or angle) when patients are unable to maintain their positions. The previous articles on ML for wound assessment were usually trained from a limited dataset. Although the performances in papers are satisfactory, the real applications show the opposite. Compared with traditional ML, deep learning (DL) can comprehend all unfavored factors rather than exclude them in order to get better results. For example, the luminance component was usually eliminated to minimize the effect of lighting. Although the elimination of the luminance component will produce better validation, it makes the model difficult to apply in actual clinical conditions. However, training DL models requires plenty of labeled images. A systemic processing and labeling method to set up datasets is most crucial. Different types of wounds need to be described by various outputs. For example, the assessment of acute burn wounds is to calculate the percentage of total body surface (%TBSA) burned and the deep burn area requiring debridement or graft, whereas the evaluations of chronic ulcers focus on types of tissues and the absolute size of wounds. All the parameters are the extension of two tasks: wound segmentation and tissue classification/segmentation. Wound segmentation is to differentiate all areas of wounds from normal skin. The results are transferred to the size of the wounds. Tissue classification is to segment different tissues inside and peri-wound. The granulation on pressure ulcers is a type of tissue, while the deep burn without perfusion is also a type of tissue inside the whole burn area. The two concepts can be applied to any wound, from acute burns to chronic ulcers. In this study, the boundary-based labeling method is used for wound edge and peri-wound tissues labeling. The region-based labeled method by superpixel segmentation pre-processing is applied for tissue labeling inside wounds. Both acute burn wounds and chronic ulcers can build the high quality and standard datasets by the approach. Several DL models training from these datasets have decent results for wound segmentation and tissue classifications.