開發基於光學同調斷層掃描皮膚影像之自動化分層以及光學特性分析演算法

Abstract

現今在評估皮膚疾病時，醫生通常需透過皮膚鏡來放大檢查可疑之病灶區域，故皮膚鏡常被作為第一線的檢查儀器。然而，經由皮膚鏡診斷的準確性需高度仰賴醫生的經驗。近來，不少先進生醫光學影像技術被提出，像是反射式共軛焦顯微術(reflectance confocal microscopy, RCM)和光學同調斷層掃描術(optical coherence tomography, OCT)。儘管RCM可達到細胞級的皮膚結構影像，但視野(field of view, FOV)和成像深度皆受到限制。對於OCT而言，雖然OCT可提供三維皮膚結構影像，但目前大多數的OCT系統不論是成像分辨率或是成像深度都有所限制。除此之外、目前市面上 RCM 或 OCT 系統體積皆較為龐大。除前述硬體限制之外、前述兩項技術或市面上的系統對於提供定量化影像分析則相當具有挑戰性。

在本篇論文中，我們將致力於突破前述系統或技術在硬體及軟體層面的限制。在硬體方面，我們使用芯聖科技(OPXION Technology, Inc.)所開發之手持式3D皮膚影像掃描儀，其基於OCT技術並具有三維皮膚組織成像特點，與市面上OCT相比，體積小且輕量化。至於軟體方面，我們根據手持式OCT系統中獲取的皮膚OCT影像和OCT影像中局部光學衰減係數(optical attenuation coefficient, OAC)，開發了全自動化OCT皮膚影像分析演算法。此演算法可量化皮膚狀況並計算多項皮膚參數，包括表皮厚度、皮膚粗糙度及表皮和真皮的紋理。

為了提升全自動化皮膚影像分割演算法中的準確率與處理速度，我們開發以U-net模型為基底的深度學習演算法。在深度學習演算法當中，我們使用局部光學衰減係數計算的分割結果並且進一步透過人為檢視針對不正確的分割結果進行手動校正，且以此一校正過的分割結果作為模型開發的基礎。初步的研究結果顯示可有效改善分割之準確率和處理速度。

此外，我們探討市售肌膚水份檢測儀與OCT之關聯性，使用市售保濕面膜進行試驗，比較受試者使用前後的皮膚參數與OCT影像之差異。在未來的研究中，我們將致力整合演算法於客製化OCT使用者介面(Graphical User Interface, GUI)，並對深度學習模型進行優化，增加更多不同部位之人體皮膚影像，以期在面對各式皮膚型態時也能有效準確分割。我們期望本碩士論文的研究成果能成為未來OCT跨足醫美的基石，尤其是肌膚護理領域。再通過與臨床醫學及醫美診所的團隊們多方合作，以期最大化本研究成果。

Currently, for the assessment of skin diseases, physicians usually rely on dermoscopy to provide magnified visualization of the suspected regions as the first-in-line screen tool. However, the diagnostic accuracy of dermoscopy often relies on physicians’ experience. Recently, the emergence of advanced biomedical optical imaging technologies, such as reflectance confocal microscopy (RCM) and optical coherence tomography (OCT). Although RCM can provide cellular-level imaging of the skin architecture, both the field of view (FOV) and the imaging depth are limited. On the other hand, although OCT can provide volumetric images of the skin architectures, most of the currently available OCT system, either the imaging resolution or the imaging depth, is limited. Also, the size of either the commercially available RCM or OCT system is relatively bulky. Furthermore, in addition to above existing hardware limitations, it is relatively challenging to provide quantitative analysis with either system or technologies.

In this master thesis work, we aim to address the above limitations from both the hardware and software perspectives. For the hardware side, we used the handheld 3D skin viewer system developed by OPXION Technology, Inc., which can provide volumetric imaging of the skin tissue architectures based on the OCT technology but exhibits a smaller size and weight compared to the existing OCT systems. As for the software side, we have developed a fully automatic analysis algorithm based on the skin OCT images acquired using the handheld OCT system and the local attenuation coefficient in the OCT images. Various parameters quantifying the skin conditions, including the thickness of the epidermis, skin roughness, and texture of the epidermis and dermis, can be computed with the developed algorithm.

In addition, in order to improve the accuracy and computation speed of the above skin image segmentation algorithm, we have developed a deep-learning algorithm based on the U-net model. In this deep-learning algorithm, we used the segmentation results computed based on the local attenuation coefficient where incorrect segmentation results have been manually corrected as the basis for the development of the deep-learning model. The preliminary results have shown an improved performance in both segmentation accuracy and computation time.

Furthermore, we also investigated the correlation between the commercial skin moisture tester and OCT and compared the differences in the skin OCT parameters for participants before and after applying the farcical mask. In the future, we will focus on integrating the algorithm into the OCT customized graphical user interface (GUI) and optimizing the deep learning model, for example, by including more skin images collected from different parts of the human skin. Therefore, the segmented results can be more effectively and accurately predicted in the face of various skin types. We believe the results of this master thesis work set the footstone for future translation of OCT into aesthetic medicine, particularly the skincare regime. In future research, through the cooperation with clinical medicine and medical aesthetic clinics teams, we expect our research can produce the best results.