皮膚的層次以及細胞核的影像分割無論在診斷還是電腦分析,都是病理學上重要的資訊及任務。本論文中,我們設計了一個可以從人類皮膚的三維光學同調斷層掃描(OCT)中獲取細胞核體機率分佈的架構,再利用得來的細胞核資訊,進一步訓練出分割細胞核的模型,以用來輔助影像轉換,該影像轉換建立在循環生成對抗網路並能將人類皮膚的光學同調斷層掃描影像轉換成蘇木精與伊紅(H E)染色。在測試影像集上,角質層的下邊界與表皮層的下邊界於影像轉換前後的位置誤差分別為±0.73μm和±4.53μm,而在加入細胞核位置資訊後,細胞核於影像轉換前後的索倫森-骰子係數從0.536上升至0.586。
Skin layers and nuclei segmentation is a crucial task for computational pathology applications and it assists the diagnosis of skin cancers. In this research, we proposed a framework to achieve nuclei annotation of in vivo optical coherence tomography (OCT) image of three-dimensional human skin structure. Furthermore, we trained a segmentation model with the pseudo annotation of nuclei, to facilitate CycleGAN-based image conversion of human skin from OCT images to hematoxylin and eosin (H E) stain. Input OCT images and the output H E-like images match well in terms of the stratum corneum lower boundaries and dermal-epidermal junction with errors of ±0.73 μm and ±4.53μm, respectively. Dice similarity coefficients of the nuclei consistence are 0.427 and 0.536 before and after adding nuclear info.