Detection of shredded split in breast cancer cells (2)

The method used in mitosis detection in breast cancer histology images with deep neural networks has achieved good results and won the first place in the icpr2012 detection and split competition.

Principle:

We used a deep Neural Network (dnn) to detect the split of breast cancer cells. Let's learn about the principles of dnn. This article uses a sliding window to sample the original image (using a small rectangle), and then uses dnn to detect the image.

Process:

Our goal is to find such a set. Each element represents the center of a split core.

Using the given labeled samples, we train a classifier. each pixel in the sample has a tag (split with filaments or not split with filaments ), pixels close to the split-core center are also split. For each pixel P to be predicted, we use a p-centered Sliding Window for sampling, and then use dnn for prediction.

The methods proposed in this article mainly include: 1. Understanding dnn; 2. Understanding the training and testing process.

Dnn belongs to the scope of machine learning. Let's learn it by ourselves. Let's talk about the second point of this article.

If you understand the principles of dnn, you should understand the training process. Note that here, the author thinks there are too few positive samples, we can use any Rotating sampling rectangle to obtain more positive samples and solve the rotation immutability during the detection process.

Let's take a look at how to process an image to be tested:

For an image to be processed, we use dnn on all sampling windows. the center of each window must be within the image boundary, in this way, a probability graph M can be generated (giving each pixel the probability of approaching the split-core center), and then convolution with a circular kernel function with a radius of D pixels, A probability map MF after smooth processing can be generated. The local maximum value of MF is considered as the center of each shredded Split Core.

To put it simply, as shown in the figure below:

Source image I (using dnn) --> probability graph M (Convolution with the circular kernel function with a radius of D) --> smoothing probability map MF --> the local maximum value of MF is the center of the split-core.

As we mentioned earlier, our goal is to get it. How can we get it?

1. First, set D to null;

2. iteration steps:

A. Assume that PM is a local maximum point in MF and its position is classified into D;

B. Set MF (p) to 0 if the distance between P and PM is less than the given 2D value.

3. Until all MF (PI) values are less than the threshold T.