以特定色彩元素和特徵集組合分析肝臟組織切片來預測老鼠之年齡

Abstract

細胞老化被研究 40 年餘，已有許多文獻指出在分子層面上，年輕和老化細胞的差異。在形態表現上，雖然有觀察到細胞老化所造成的改變，例如組織細胞核上出現形狀的缺陷，但尚不能給予數值化，而進一步實現更重要的問題，預測未知個體的老化程度。 為了能夠準確對於未知個體預測，本研究以老鼠肝臟組織切片影像找出適合用於分類四類年齡 (一個月、六個月、十六個月、二十四個月) 的特徵集和各種色彩空間的元素組合，即八種特徵集和 23 個色彩元素進行留一個體測試。結果顯示 GLCM 和 Haralick 此兩組特徵集組合效果最好，並將此兩種特徵集結合在一起之後，發現 CMYK 色彩模型的 C 元素準確率可達 76.19%。但當僅使用單一通道時僅有 66.67%。結果明確指出和傳統單一通道方式相比，組合效果能夠有效提升準確率。 本研究除了提高留一個體測試的準確率，在不同年齡的蘇木精-伊紅染色肝臟切片影像上，找出是因為在 C 元素下使用外觀紋路相關特徵，而對老鼠組織切片分類年齡有較高的準確率。因為蘇木精-伊紅染色會使細胞核呈現藍色，因此可推知不同的老化程度其細胞核在外觀紋路上會有所區別。另外本研究首次提出了「多色彩空間通道組合」運用於影像分類問題，此方式能夠搭配特徵選取提出影像分類上的差異，而此強力工具也能夠應用於其他影像分類問題上。

It has been 40 years for studying cell senescence. There are many literatures which point the differences between young and senescent cells in molecular level. In morphological way, the changes on surface of senescent cells are observed. But it is hard to represent the variation with values. Consequently, predicting the hierarchy of senescence that is more important problem remains unaccomplished. In this study, the GLCM and Haralick feature sets which had been used for describing tissues are applied to predict unknown class of individuals. This study applied different combination of feature sets and color components for classifying four ages (1 month, 6 month, 16 month, 24 month) with images of mouse liver tissue biopsy. With 23 components of color spaces, leave-one-subject-out test gives an ensemble result 76.19%. On the other hand, using RGB-based gray scale component only reach to 66.67%. Comparing to single component, these results reveal that multi-component of color spaces ensemble way increases the prediction accuracy. Except the higher classification accuracy with leave-one-subject-out, we prove that the C component of CMYK color space with texture feature sets improves the classification accuracy between different ages of hematoxylin and eosin stain images. Because the defect of cell nuclei could be expressed in blue color with hematoxylin and eosin stain, the level of senescence would be identified on the morphological shape and texture of cell nuclei. On the other hand, this study is the first one to show the combination of multi-component ensemble for images classification. The method is powerful for identifying the difference between images with feature selection, and for other problems of image classification.