A research team led by Professor Park Sang-Hyun from the Department of Robotics and Mechatronics Engineering (also in charge of the Artificial Intelligence specialty) of DGIST (Chairman Kuk Yang) developed a learning model unsupervised deep that can accurately show presence and location. of cancer in pathological images based only on data where cancer is present. Existing deep learning models needed to build a dataset, on which the location of the cancer was accurately drawn, to specify the site of the cancer. The deep learning model developed in this study improved the efficiency and is expected to make a significant contribution to the relevant research field.
In general, it is necessary to accurately mark the location of the cancer site to solve the zoning problems indicated by the cancer location information, which is time-consuming and thus increases the cost.
To solve this problem, the weakly supervised learning model that zones cancer sites with only rough data such as “whether cancer in the image is present or not” is under active study. However, it would significantly degrade the performance if the existing weakly supervised learning model is applied to a large dataset of pathological images where the size of an image is as large as a few gigabytes. To solve this problem, researchers tried to improve the performance by dividing the pathological image into patches, but the divided patches lose the correlation between the location information and the divided data, which means that there is a limit to using the whole information available.
In response, Professor Park Sang-Hyun’s research team discovered a segmentation technique to the cancer site based solely on the learned data indicating the presence of cancer per slide. The team developed a pathological image compression technology that first teaches the network to effectively extract meaningful features from patches using unsupervised contrastive learning and uses it to detect salient features while keeping information from each location to reduce the image size and maintain correlation between patches. Later, the team developed a model that can find the region that is most likely to have cancer from the compressed pathology images using a class activation map and a zone of all regions that are most likely to have cancer. have cancer from whole pathology images using a pixel correlation module (PCM).
The recently developed deep learning model showed a Dice Similarity Coefficient (DSC) score of up to 81 – 84 with only the slide-level cancer-labeled learning data in the cancer zoning problem. It significantly outperformed previously proposed patch-level methods or other unsupervised learning techniques (DSC score: 20 – 70).
“The model developed through this study has greatly improved the performance of weakly supervised learning of pathological images and is expected to contribute to improving the efficiency of various studies that require analysis of pathological images,” he said Professor Park Sang-Hyeon Park who added, “If we can further improve the related technology in the future, it will be possible to use it universally for various medical image zoning problems.”
Meanwhile, the results of this study were recognized for their excellence and published in MediIA (Medical Image Analysis Journal), an international academic journal with the highest authority in the field of medical image analysis.
Source:
DGIST (Daegu Gyeongbuk Institute of Science and Technology)
Journal reference:
10.1016/j.media.2022.102482