Limits...
Unsupervised segmentation of low-contrast multichannel images: discrimination of tissue components in microscopic images of unstained specimens.

Kopriva I, Popović Hadžija M, Hadžija M, Aralica G - Sci Rep (2015)

Bottom Line: Such images are very hard to segment.The method is a proof-of-concept validated on the unsupervised segmentation of color images of unstained specimens, in which case the tissue components appear colorless when viewed under the light microscope.The proposed method can potentially be applied to the segmentation of low-contrast multichannel images with high spatial resolution that arise in other imaging modalities.

View Article: PubMed Central - PubMed

Affiliation: Division of Laser and Atomic Research and Development, Ruđer Bošković Institute, Bijenička cesta 54, 10002 Zagreb, Croatia.

ABSTRACT
Low-contrast images, such as color microscopic images of unstained histological specimens, are composed of objects with highly correlated spectral profiles. Such images are very hard to segment. Here, we present a method that nonlinearly maps low-contrast color image into an image with an increased number of non-physical channels and a decreased correlation between spectral profiles. The method is a proof-of-concept validated on the unsupervised segmentation of color images of unstained specimens, in which case the tissue components appear colorless when viewed under the light microscope. Specimens of human hepatocellular carcinoma, human liver with metastasis from colon and gastric cancer and mouse fatty liver were used for validation. The average correlation between the spectral profiles of the tissue components was greater than 0.9985, and the worst case correlation was greater than 0.9997. The proposed method can potentially be applied to the segmentation of low-contrast multichannel images with high spatial resolution that arise in other imaging modalities.

No MeSH data available.


Related in: MedlinePlus

Human liver with hepatocellular carcinoma.(a) red-green-blue (RGB) color microscopic image of unstained specimen. (b) “ground truth – different slides” RGB color microscopic image of the specimen stained by Hep Par. (c) color-coded (digitally stained) image of the segmentation result obtained by the EKM-NMF_L0 algorithm (Gaussian kernel with variance = 0.1 and D = 50): blue: hepatocellular carcinoma, red: blood vessel, green: tumor fibrotic capsule. (d) color-coded (digitally stained) image of the segmentation result obtained by the NMF_L0 algorithm. (e) RGB color microscopic image of the specimen (a) stained subsequently by H&E. (f) average mutual coherence for the original matrix of the spectral profiles and matrices in the EKM-induced space. (g) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in RGB (1-2-3) color space, μ(A) > 0.9999, μ_average(A) = 0.9985. (h) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in non-physical color space induced by EKM (D = 50, Gaussian kernel variance=0.1), μ(B) = 0.9760. μ_average(B) = 0.8937.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4477329&req=5

f2: Human liver with hepatocellular carcinoma.(a) red-green-blue (RGB) color microscopic image of unstained specimen. (b) “ground truth – different slides” RGB color microscopic image of the specimen stained by Hep Par. (c) color-coded (digitally stained) image of the segmentation result obtained by the EKM-NMF_L0 algorithm (Gaussian kernel with variance = 0.1 and D = 50): blue: hepatocellular carcinoma, red: blood vessel, green: tumor fibrotic capsule. (d) color-coded (digitally stained) image of the segmentation result obtained by the NMF_L0 algorithm. (e) RGB color microscopic image of the specimen (a) stained subsequently by H&E. (f) average mutual coherence for the original matrix of the spectral profiles and matrices in the EKM-induced space. (g) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in RGB (1-2-3) color space, μ(A) > 0.9999, μ_average(A) = 0.9985. (h) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in non-physical color space induced by EKM (D = 50, Gaussian kernel variance=0.1), μ(B) = 0.9760. μ_average(B) = 0.8937.

Mentions: Segmentation of nontrivial images is considered one of the most difficult tasks in image processing1. Image segmentation refers to the partitioning of an image into sets of pixels (segments) corresponding to distinct objects2. Within the scope of this study, distinct objects refer to spectrally distinct tissue components present in the images of unstained specimens. It is important to distinguish between single (grayscale)- and multi-channel images. In the former case, segmentation is performed by detection of changes of intensity or texture by thresholding some type of spatial derivative of an image34567. However, images that comprise components with very similar profiles (spectral, density, and/or concentration) have very low visual contrast. For an example, if staining is not used, the spectral similarity between the tissue components present in the specimen is very high and the visual contrast is very poor, i.e., tissue components appear colorless and virtually texture-less when viewed under a light microscope. This situation occurs in the case of synthetic images (Fig. 1a), as well as in case of color microscopic images of unstained specimens of human hepatocellular carcinoma (primary liver tumor) (Fig. 2a), liver tissue with metastasis from colon cancer (Fig. 3a), and gastric cancer (Fig. 4a). Thus, when spectral vectors are plotted vs. their indices (corresponding red, green and blue colors) they are virtually parallel (Figs. 1g,2g and Figs 2a,3a,4a, S5a and S6a). Consequently, the intensity and/or texture-based segmentation methods34567 fail to segment tissue components correctly (Figures S2b and S2c). Segmentation of the color image by means of clustering in the CIE L*a*b* color space8 also fails for the same reason (Fig. 1e, S2a, S3a, S4a, S5e and S6f).


Unsupervised segmentation of low-contrast multichannel images: discrimination of tissue components in microscopic images of unstained specimens.

Kopriva I, Popović Hadžija M, Hadžija M, Aralica G - Sci Rep (2015)

Human liver with hepatocellular carcinoma.(a) red-green-blue (RGB) color microscopic image of unstained specimen. (b) “ground truth – different slides” RGB color microscopic image of the specimen stained by Hep Par. (c) color-coded (digitally stained) image of the segmentation result obtained by the EKM-NMF_L0 algorithm (Gaussian kernel with variance = 0.1 and D = 50): blue: hepatocellular carcinoma, red: blood vessel, green: tumor fibrotic capsule. (d) color-coded (digitally stained) image of the segmentation result obtained by the NMF_L0 algorithm. (e) RGB color microscopic image of the specimen (a) stained subsequently by H&E. (f) average mutual coherence for the original matrix of the spectral profiles and matrices in the EKM-induced space. (g) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in RGB (1-2-3) color space, μ(A) > 0.9999, μ_average(A) = 0.9985. (h) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in non-physical color space induced by EKM (D = 50, Gaussian kernel variance=0.1), μ(B) = 0.9760. μ_average(B) = 0.8937.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4477329&req=5

f2: Human liver with hepatocellular carcinoma.(a) red-green-blue (RGB) color microscopic image of unstained specimen. (b) “ground truth – different slides” RGB color microscopic image of the specimen stained by Hep Par. (c) color-coded (digitally stained) image of the segmentation result obtained by the EKM-NMF_L0 algorithm (Gaussian kernel with variance = 0.1 and D = 50): blue: hepatocellular carcinoma, red: blood vessel, green: tumor fibrotic capsule. (d) color-coded (digitally stained) image of the segmentation result obtained by the NMF_L0 algorithm. (e) RGB color microscopic image of the specimen (a) stained subsequently by H&E. (f) average mutual coherence for the original matrix of the spectral profiles and matrices in the EKM-induced space. (g) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in RGB (1-2-3) color space, μ(A) > 0.9999, μ_average(A) = 0.9985. (h) spectral responses of the tumor fibrotic capsule, blood vessel and hepatocellular carcinoma in non-physical color space induced by EKM (D = 50, Gaussian kernel variance=0.1), μ(B) = 0.9760. μ_average(B) = 0.8937.
Mentions: Segmentation of nontrivial images is considered one of the most difficult tasks in image processing1. Image segmentation refers to the partitioning of an image into sets of pixels (segments) corresponding to distinct objects2. Within the scope of this study, distinct objects refer to spectrally distinct tissue components present in the images of unstained specimens. It is important to distinguish between single (grayscale)- and multi-channel images. In the former case, segmentation is performed by detection of changes of intensity or texture by thresholding some type of spatial derivative of an image34567. However, images that comprise components with very similar profiles (spectral, density, and/or concentration) have very low visual contrast. For an example, if staining is not used, the spectral similarity between the tissue components present in the specimen is very high and the visual contrast is very poor, i.e., tissue components appear colorless and virtually texture-less when viewed under a light microscope. This situation occurs in the case of synthetic images (Fig. 1a), as well as in case of color microscopic images of unstained specimens of human hepatocellular carcinoma (primary liver tumor) (Fig. 2a), liver tissue with metastasis from colon cancer (Fig. 3a), and gastric cancer (Fig. 4a). Thus, when spectral vectors are plotted vs. their indices (corresponding red, green and blue colors) they are virtually parallel (Figs. 1g,2g and Figs 2a,3a,4a, S5a and S6a). Consequently, the intensity and/or texture-based segmentation methods34567 fail to segment tissue components correctly (Figures S2b and S2c). Segmentation of the color image by means of clustering in the CIE L*a*b* color space8 also fails for the same reason (Fig. 1e, S2a, S3a, S4a, S5e and S6f).

Bottom Line: Such images are very hard to segment.The method is a proof-of-concept validated on the unsupervised segmentation of color images of unstained specimens, in which case the tissue components appear colorless when viewed under the light microscope.The proposed method can potentially be applied to the segmentation of low-contrast multichannel images with high spatial resolution that arise in other imaging modalities.

View Article: PubMed Central - PubMed

Affiliation: Division of Laser and Atomic Research and Development, Ruđer Bošković Institute, Bijenička cesta 54, 10002 Zagreb, Croatia.

ABSTRACT
Low-contrast images, such as color microscopic images of unstained histological specimens, are composed of objects with highly correlated spectral profiles. Such images are very hard to segment. Here, we present a method that nonlinearly maps low-contrast color image into an image with an increased number of non-physical channels and a decreased correlation between spectral profiles. The method is a proof-of-concept validated on the unsupervised segmentation of color images of unstained specimens, in which case the tissue components appear colorless when viewed under the light microscope. Specimens of human hepatocellular carcinoma, human liver with metastasis from colon and gastric cancer and mouse fatty liver were used for validation. The average correlation between the spectral profiles of the tissue components was greater than 0.9985, and the worst case correlation was greater than 0.9997. The proposed method can potentially be applied to the segmentation of low-contrast multichannel images with high spatial resolution that arise in other imaging modalities.

No MeSH data available.


Related in: MedlinePlus