Denoising AutoEncoder-based Representation Learning for Multi-Task Whole Slide Image Analysis

Abstract

Listen

Translate article

Along with the advancement of artificial intelligence, there have been significant improvements in the field of whole slide images (WSI). WSI in machine learning is mainly utilized for pathological analysis, consisting of diverse tasks such as classification of normal versus tumor patches, segmentation of precise areas of potential tumor, or object detection indicating tumor sites. However, training these distinct models for each individual task is both time intensive and inefficient. Therefore, there is a high demand for developing a unified learning algorithm capable of concurrently handling multiple WSI tasks. To address the aforementioned problem, a representation learning based transfer learning method is proposed to process multiple downstream tasks including classification, segmentation, and object detection. Synthesizing two stages, the proposed method utilizes the reconstruction of images from representation learning and pretrained parameters from transfer learning method to create a more accurate and time-efficient model for analyzing WSI. Overall, the proposed method offers a better representation of WSI, which leads to enhanced accuracy in analysis and interpretation. Through extensive experiments, I have found that the proposed method outperforms previous state-of-the-art networks in various downstream tasks including classification, segmentation, and object detection. I expect the proposed method to be applied in real world scenarios with increased practicality and accuracy.