Improving Medical Image Analysis with CNNs: Lunit Patent Analysis, Part 2

Following an earlier analysis of Lunit's medical-imaging patents, this article examines another CNN-based approach: analyzing cell images to identify the number and location of mitotic cells.

The previous article examined training a CNN to improve detection accuracy. This article examines how a CNN identifies the number and locations of mitotic cells.

Click the image to view the granted patent.

A cell image is supplied to the first convolutional layer to extract a feature map. 

The second convolutional layer processes that feature map and produces the first class activation map.This map first class activation map.represents features associated with mitotic cells.

first class activation map.It is then first class activation map.processed through the output stage using global pooling, such as max or average pooling. first class activation map.The resulting second class activation mapis used in the following step.

second class activation mapThe map is then enlarged second class activation mapbecause convolutional processing has reduced its spatial dimensions relative to the input image. second class activation mapTo locate the mitotic cells in the original image, second class activation mapthe activation map is resized to match the input's dimensions. This is called the third class activation mapin the description.

third class activation mapThe map is then enlarged third class activation mapRegions exceeding a threshold or containing a local peak are identified; the illustration shows them in darker colors. third class activation mapThose regions indicate candidate mitotic-cell locations in the original image. Counting the regions estimates the number of mitotic cells.  

Although the underlying CNN operations are familiar, the patent describes their use for locating and counting mitotic cells.That specific application is central to the analysis.