3.1 Feature extraction

We have chosen a variety of features from literature review to explore. [1] These can be divided into two categories, binary and grayscale, which refer to the types of images that the code operates on. The binary features are based on morphological properties including cell and nucleus area and perimeter. The grayscale features describe the texture and contrast of the cells. We implemented a total of 12 features, 5 binary and 7 grayscale. All the features below were calculated for both the cell and the nucleus, except for the ratio of nucleus area to cell area.

We used 150 test images in order to extract features for our matrices. Again, our images were 250 x 250 pixels which we believed to be large enough to minimize errors from segmentation.

Binary features

Ratio of nucleus to cell area

We calculate area by summing all of the white pixels in an image into a scalar number, and compute the ratio by dividing these numbers.

Ratio of area to perimeter

Circularity

Compactness

Grayscale features

The grayscale features are based on the gray-level co-occurrence matrix (GLCM), which can be used to extract statistical measures of texture. The GLCM is a matrix with elements p(i,j) that are equal to the number of times in the image a pixel with grayscale intensity i appears adjacent to a pixel with grayscale intensity level j. [3]

Contrast

Homogeneity

Entropy