AI Solution: DEEP:PATH-CRC-01
Core Technology: Automatic region extraction technology of abnormal tissue in colon cancer pathology image
Colorectal Cancer is a malignant tumor that occurs in the colon and rectum. Colorectal cancer has no symptoms at the beginning, and when symptoms appear, it is often quite advanced. Therefore, if colon cancer is detected and treated at an early stage, the treatment results are very good. Therefore, it is very effective to find the polyp in the adenoma stage through examination and remove it with an endoscope. Tissues removed during endoscopy are diagnosed as benign or malignant through biopsy. There are several methods for early diagnosis of colon cancer, such as CT, blood tests, and molecular tests, but the method of finally examining the tissue obtained through an endoscope is the most accurate. A biopsy is a diagnostic method in which tissue obtained through an endoscope is sliced, stained, and observed through a microscope.
The structure of the large intestine tissue in the human body is shown in the picture above. When observed with a pathology microscope, it can be classified into five broad categories, which can only be diagnosed by a pathologist. At this time, pathologic images are images used as the final stage of diagnosing diseases, unlike other medical images used for initial diagnosis, so if you misdiagnose, you have a lot of damage and a lot of responsibility compared to other medical images. In addition, the time it takes for a pathologist to read a single image can take several hours. Therefore, in order to maximize the diagnosis rate and time efficiency of pathologists, we have developed artificial intelligence software for colorectal cancer pathology that extracts and visualizes abnormal tissue areas.
AI Core Technology
Pathology image data, which is a microscopic image, has a very large size unlike other data. However, due to the limitations of computer capacity, smaller images are required to train artificial intelligence. At this time, if the original image is simply reduced to a small size, the resolution is lost and the diagnostic accuracy is lowered.If the original image is simply extracted and learned, the region of interest that can be determined once learned is reduced, which also increases noise. This results in lower accuracy.
Therefore, we have devised a method of compressing the image using principal component analysis and wavelet transformation as above, so that there is no damage to the resolution and has a large area of interest. At this time, principal component analysis is a method that efficiently compresses the correlation of the three color channels (RGB-channels) constituting an image in a low-dimensional manner. Through this, the meaningful components contained in each channel can be compressed into a black-and-white image and efficiently analyzed by removing the meaningless background area without tissue. In addition, wavelet transformation is a method of reducing the resolution without loss by using a signal whose frequency component changes over time. This technique is also used in the popular JPEG 2000 compression method as a method of compressing general digital images.
These compressed patch images are trained and analyzed using the UNet++ architecture with InceptionResNet backbone to obtain results. InceptionResNet is a classification model that combines the features of InceptionNet and ResNet, which won the 2014 and 2015 championships in ILSVRC (ImageNet Image Recognition Competition). InceptionNet is a neural network model released by Google that reduces the size of the existing relatively large convolution filter to a small form and configures it into several, so that operations can be divided into small forms. In addition, when training a deep neural network, ResNet, which solved the phenomenon of overfitting to the training data, was attached.
The UNet++ architecture is a model that shows more accurate performance by using the skip-connection part of the existing UNet for the purpose of analyzing and delivering features rather than simply delivering the original image. By combining these three models, normal and abnormal areas of the colon tissue were detected to maximize the performance of histopathology examination. The results derived from the neural network are combined with the results of analyzing each feature through inverse wavelet transformation, and the resulting patch is combined with the original image size and overlaid on the original image.