Detection of split in breast cancer cells (1)

ICPR and miccai have been paying close attention to the split detection of breast cancer cells in recent years. I have also studied it, although I don't know what I can do in the end.

Today I read this article "automatic mitosis Detection Based on exclusive independent component analysis", published on Pattern Recognition (ICPR), 2012 21st International Conference, the main research is to detect the split of breast cancer cells.

Abstract:

We have adopted a new method called exclusive independent component analysis (xica), which is an extension of ICA. We know that ICA is called independent component analysis, so xica is easy to understand, it mainly extracts different components from each other in the training set, rather than the independent components of a single category.

Xica:

This algorithm is the key of this Article. Let's first understand it.

This algorithm attempts to find such independent components in a training set with two categories: it is closely related to one of the classes, but not very representative of the other, it can also be said to look for some specific components.

Generally, the ICA algorithm can be written in the following format:

X =.

X = (x1, x2, x3 ..... XN) is some of our observed vectors. We can think of them as feature vectors, S = (S1, S2, s3 ...... SN) is an independent potential variable vector. A is an unknown constant matrix (which can be called a hybrid matrix or a diversity). Our goal is to obtain a through only X vectors.

Two training modes are provided,

E {.} is the expectation of the input value, but the corresponding weight matrix. The method is as follows:

The formula above can calculate W. As for how to calculate G {.}, I have to go to the paper A. hyv? arinen and P. O. Hoyer. emergence of phase and.
Shift Invariant Features by decomposition of neutral images into independent feature subspaces. Neural computing, 12 (7): 1705-1720,200 0.

The following iteration is used to normalize the two matrices.

W can produce independent components of positive sample X, but cannot produce independent components of negative samples, and vice versa.

After obtaining a and S, the following test is simple,

(Another method of the above formula) to minimize

After finding S, perform the following steps: X is the new test set and calculates the difference between it and AKs, select K corresponding to the minimum value as the category of X (k is the number of classes), as shown below:

Of course, we do not calculate all the pixel values. First, we should make a selection to obtain the candidate set.

Using a simple cell nucleus detection algorithm: first, the image is converted into a grayscale image, and then Gaussian smoothing is used to tune it to appropriate parameters. Most of the cells are protruding, we select some local maximum values from these convex points to join the candidate set.

The next task is to use the latest algorithms in these candidate sets.