應用深度學習定量和定性分析細胞解析度光學同調斷層掃描影像中的人體皮膚結構和病變

Abstract

醫學影像分析不論是在醫學、工業還是學術界一向為一個錯綜複雜的問題。以定量和定性地解釋醫學影像需要長期的經驗和專業知識的累積。以光學同調斷層掃描（Optical coherence tomography, OCT）影像而言，其具備非侵入性和高速診斷的優勢，甚至近幾年已達到細胞級解析度的水準。然而在OCT影像中，對人體皮膚結構和病變的研究仍然有限，所以本論文的目標是通過開發一系列深度學習演算法和模型架構，來探索微米解析度等級的人體皮膚OCT影像。

為了能夠更系統性的分析和理解，本論文總共劃分成四個大項。第一大項為「定量理解」。一開始將開發二維影像分割模型，分割人類皮膚層和細胞核，並探討二維模型的侷限性，之後則進一步開發非監督式影像去雜訊模型先改善三維影像的品質，再發展半監督式三維影像分割模型解決二維模型的限制，並分析人類角質細胞核的大小。第二大項為「定性理解」。此項目中將透過開發生成式模型轉換OCT和醫學中常見的蘇木精與伊紅(Hematoxylin and eosin, H&E)染色切片影像，來理解人類皮膚組織中兩種影像的對應性，而這個部分會先以非監督式學習的方法來初步了解轉換上的效果和限制，之後將進一步利用影像標註來增加轉換的準確性。

基於前兩大項中的知識背景，接下來將探討具有疾病的人類皮膚OCT影像。而第三大項為「定性解釋」。此項目將開發一系列影像辨識模型來區別不同的疾病，並透過可解釋性人工智慧框架Grad-CAM用以辨別各個疾病的病兆，其中將先以自監督式學習開發常見皮膚疾病的辨識模型，接著以元學習開發數據量非常少的罕見疾病辨識模型，最後將針對長尾資料分布情境，開發一套可辨識健康狀態和13種皮膚疾病的模型。第四大項為「定量解釋」。在第三大項中，由於神經網絡可解釋性上的限制，無法進一步解釋疾病的嚴重程度和不同情境定量上的判斷依據，所以此項目將先以多任務學習強化二維影像分割模型在疾病影像的適應性，以及賦予模型自動辨識出黑色素分布的能力，並以此分割結果製作一系列的指標和特徵來開發可定量解釋的機器學習模型。

對於本論文的主要貢獻不僅在OCT影像上分析人體皮膚提供定量和定性的視角，並更進一步奠定了可信賴人工智慧和精準醫療診斷的基礎。

Medical image analysis is a convoluted and involuted problem whether in the field of medicine, industry, or academia. Quantitatively and qualitatively interpreting the medical data invariably and inevitably necessitate the long-term experience and long-standing expertise. While the progress of the cellular-resolution optical coherence tomography (OCT) present the opportunity for non-invasive and high-speed diagnosis and providing histopathological-level information, limited amounts of research have explored the human skin structures and lesions in such images. The advancement of the deep learning paves the way for systematically and automatically analyzing the medical images. Therefore, the objective of this thesis is exploring the human skin cellular-resolution OCT imaging through developing a series of deep learning approaches.

In order to facilitate a more holistic analysis, this thesis is divided into four major components. The first component is "Quantitative Comprehension." Initially, a 2D image segmentation model will be developed to segment healthy human skin layers and cell nuclei, investigating the limitations of the two-dimensional model. Subsequently, an unsupervised image denoising model will be developed to improve the quality of 3D volume, followed by the development of a semi-supervised 3D image segmentation model to address the limitations of the two-dimensional model, analyzing the size of human keratinocytes cell nuclei. The second component is "Qualitative Comprehension." In this aspect, a generative model will be developed to transform OCT images into commonly used hematoxylin and eosin (H&E) stained slides, in order to understand the correspondence between the two types of images in human skin tissue. This part will initially employ unsupervised learning methods to gain preliminary insights into the effectiveness and limitations of the transformation. Subsequently, image annotation will be utilized to enhance the accuracy of the transformation.

Based on the knowledge background from the previous two components, the subsequent exploration will focus on unhealthy human skin. The third component is "Qualitative Interpretation." This involves developing a series of image recognition models to differentiate various diseases and employing the explainable artificial intelligence framework, Grad-CAM, to identify disease-specific signs. Initially, a self-supervised learning approach will be used to develop an image recognition model for common skin diseases. Subsequently, meta-learning techniques will be utilized to develop recognition models for scarce diseases with limited data. Ultimately, the model for the long-tail data distribution is developed to form a comprehensive system capable of recognizing both healthy conditions and 13 skin diseases. The fourth component is "Quantitative Interpretation." Due to the limitations in the interpretability of neural networks, further quantitative assessments of disease severity and context-dependent judgments cannot be made. Consequently, this component will begin with multi-task learning to enhance the adaptability of the segmentation model in disease images and identify melanin distribution. Based on the segmentation results, a set of indicators and features will be generated to develop an interpretable machine learning model.

The main contribution of this thesis not only provides the quantitative and qualitative perspective for analyzing human skin but also lays the first stone toward trustworthy artificial intelligence and accurate medical diagnosis.