利用深度學習方法從光學同調斷層掃描資料找尋真皮表皮交界處

Abstract

真皮表皮交界處(dermal epidermal junction, DEJ)是皮膚組織中重要的位置，許多皮膚癌都是從DEJ開始發展，所以在皮膚組織中尋找DEJ在輔助醫療診斷系統中是一件重要的工作。本篇論文使用光學同調斷層掃描(optical coherence tomography, OCT)所獲得之三維皮膚資料，並以深度學習方法來學習表皮與真皮層之特徵來尋找DEJ所在位置，此外，我們利用三維資料的連續性來進行前處理及後處理，讓得到的結果更加符合真實情境。實驗結果證明我們的方法比以前的方法能更加準確地找出DEJ位置，從分析中也可以看出為何我們的方法能夠成功找出DEJ。

Recently, full-field optical coherence tomography (OCT) has been developed and can get three-dimensional (3D) OCT data of human skin to achieve early diagnosis of skin cancer. In the dermatological applications of full-field OCT, dermal epidermal junction (DEJ), where melanomas and basal cell carcinomas originate, detection is an essential step for cancer diagnosis. Therefore, finding DEJ in 3D OCT data becomes an important issue for computer-aided diagnosis. However, most existing DEJ detection methods do not consider the relationship between neighboring frames. In this thesis, we proposed a novel method to find DEJ in 3D OCT data. A notable feature of our method is that it utilizes continuity of 3D data to refine the training data and train a multi-directional deep convolutional neural network (DCNN). In this way, we can eliminate noise in the training data and generate resulting DEJ with continuous surface, which follows the property of human skin. Besides, a subjective test is performed to show that the refined training data meet doctors’ standard. Finally, we evaluate our method by different metrics. The experimental results show that our method can get about 5 μm in mean error, which is significant improvement in DEJ localization.