利用深度學習演算法分析具細胞解析度的人類皮膚光學同調斷層掃描影像

Analysis of the Human Skin Tomographic Cellular-Resolution Images Utilizing the Deep Learning Algorithm

張淑雯(Shu-Wen Chang)

指導教授 ： 黃升龍

國立臺灣大學/電機資訊學院/光電工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103530

摘要

在皮膚相關的研究中，提到患有疾病的皮膚與正常皮膚的細胞核型態存在差異，若能及時提供臨床醫師皮膚細胞核的量化資訊，有機會提早診斷出皮膚是否異常。在各種皮膚疾病種類裡，基底細胞癌(Basal cell carcinoma；BCC)為常見的皮膚癌，病理學家會以組織切片的染色影像作為黃金標準來診斷，然而組織染色切片的製備流程較為耗時且複雜，若能直接從組織切片辨識出BCC，有機會加快診斷速度，減少病患忍受病痛的時間。
光學同調斷層掃描(Optical coherence tomography；OCT) 能以非侵入的方式呈現具細胞等級解析度的活體皮膚縱切面影像，也能在不將組織切片染色的情況下，提供高解析度的檢體組織影像，近年來，卷積神經網絡 (Convolutional neural network；CNN)被廣泛應用於圖像分類、目標檢測以及影像分割，其具有自動抓取影像特徵的優勢。本篇論文結合OCT成像技術與深度學習演算法，利用CNN分別從活體正常皮膚OCT縱切面影像以及檢體基底細胞癌OCT影像標記出細胞核以及基底細胞癌輪廓。針對活體正常皮膚OCT縱切面影像，細胞核分割模型mIoU達72.4%±7.8%，透過細胞核分割模型獲取細胞核標記影像，並對細胞核做量化分析，探討正常皮膚表皮層的細胞核特徵。平均細胞核大小的計算結果為20.11±4.77 μm2 (範圍：11.39 ~33.49 μm2)，另外也計算不同深度下細胞核的長短軸比，呈現正常皮膚表皮層細胞核上扁下圓的分佈趨勢。針對檢體基底細胞癌OCT影像，透過改變損失函數以及加入分類模型提升分割模型的特異性，最終patch-based準確率達87.8%±6.8%，mIoU達60.3%±10.1%。希望透過深度學習演算法的輔助，有望促進皮膚疾病的診斷與治療。

關鍵字

光學同調斷層掃描(OCT) ； 卷積神經網路(CNN) ； 影像分割 ； 活體皮膚 ； 檢體皮膚 ； 細胞核 ； 基底細胞癌

並列摘要

In dermatological researches, it is mentioned that there are some differences between the nucleus morphology of diseased skin and normal skin. Therefore, quantification of nucleus morphology could help early identification of skin abnormalities. Basal cell carcinoma (BCC) is the most common skin cancer. For pathologists, histopathological staining is considered as the gold standard for skin diagnosis. However, the staining of the excised skin tissue is a time-consuming and complicated process. Hence, it is possible to speed up the diagnosis and reduce the time for patients to endure the pain if BCC can be identified directly from the tissue section.
Optical coherence tomography (OCT) enables non-invasive imaging of in vivo skin tissue with cellular resolution. In addition, it can also provide high-resolution ex vivo skin tissue images without staining the tissue section. In recent years, the applications of convolutional neural networks (CNN) have proved the powerful performance on image classification, object detection, and segmentation. CNN has the advantage of automatic feature extraction. This thesis combines OCT imaging technology and deep learning algorithms to segment nucleus and BCC respectively from in vivo human normal skin OCT images and ex vivo BCC OCT images. For in vivo human normal skin OCT images, the mIoU of the nuclei segmentation model achieves 72.4%±7.8%. The results of the nuclei segmentation are used to explore the nuclei features in the epidermis of the normal skin. The calculation result of the mean nucleus size is 20.11±4.77 μm2 (range: 11.39 ~33.49 μm2). In addition, the ratio of the major and minor axis of the nucleus at different depths was calculated, showing a tendency that the nucleus in the upper epidermis are more oblate. For ex vivo BCC OCT images, by changing the loss function and adding the classification model to refine the specificity of the BCC segmentation model. The patch-based accuracy ultimately achieves 87.8%±6.8% and the mIoU achieves 60.3%±10.1%. The assistance of deep learning algorithms is expected to promote the diagnosis and treatment of skin diseases.

並列關鍵字

Optical coherence tomography ； Convolutional neural network ； Segmentation ； In vivo skin ； Ex vivo skin ； Nucleus ； Basal cell carcinoma

參考文獻

Y. J. Wang, Y. K. Huang, J. Y. Wang, and Y. H. Wu, ‘‘In vivo characterization of large cell acanthoma by cellular resolution optical coherent tomography,’’ Photodiagn. Photodyn. Ther. 26, 199-202 (2019)

M. Ulrich, M. Klemp, M. V. Darvin, K. Konig, J. Lademann and M. C. Meinke, ‘‘In vivo detection of basal cell carcinoma: comparison of a reflectance confocal microscope and a multiphoton tomograph,’’ J. Biomed. Opt. 18(6), 061229 (2013)

S. Astner, S. D. N. Otberg, H. R. E. Stockflesh, and J. Lademann, ‘‘Clinical applicability of in vivo fluorescence confocal microscopy for noninvasive diagnosis and therapeutic monitoring of nonmelanoma skin cancer,’’ J. Biomed. Opt. 13(1), 014003 (2008)

J. Paoli, M. Smedh, A. Wennberg, and M. B. Ericson, ‘‘Multiphoton Laser Scanning Microscopy on Non-Melanoma Skin Cancer: Morphologic Features for Future Non-Invasive Diagnostics,’’ J. Invest. Dermatol. 128, 1248-1255 (2008)

T. R. Humphreys, A. Nemeth, S. Mccrevey, S. C. Baer, and L. H Goldberg, “A pilot study comparing toluidine blue and hematoxylin and eosin staining of basal cell and squamous cell carcinoma during Mohs surgery,” Dermatol. Surgery 22(8), 693-697 (1996)