Donglin Di

B.A.

Tsinghua University

Research Assistant

NLST (Learn more about this study)

NLST-542

Aug 6, 2019

Transformer on Survival Analysis on Whole Slide Pathological Images

We propose to utilize the NLST data (CT Images and Pathology Images) to examine lung cancer disease and survival analysis (i.e., given some pathology images and then predict the possibility of the patient living 5 or 10 more years).
In order to do that, we survey a few research works on this task, like CNNs, or GCNs.
We want to apply NLST dataset to run and get the baseline results.
And after which, we propose to design a type of hypergraph CNN to improve the performance of the model by combining different but similar patients cases.

In some details, firstly, we will try to find a reasonable way to sample the WSI into several patches of images like dumping margin areas or some other ways.
Secondly, the model will extract base features from CT images or pathology images by passing through some pre-trained deep CNN network.
Thirdly, we will experiment on our model in predicting the level of health or analyzing the patient's survival, etc.

1. Screening sample patch selection model.
- Motivation: the size of each pathology image is huge, like 0.7~1.2 GB, which is not able to load whole images information into model basically. Therefore, this is the first and foremost stage to improve.
- Hypothesis: we propose to adopt the ``attention mechanism'' to detect and generate regions of key sample patches. And visualize the result of this stage output to testify.

2. Regression and prediction of the possibility of survival analysis.
- Motivation: given a pathology image, it is meaningful to predict the possibility that this patient will survive for another 5 or 10 years. And it could improve the quality of patients' lives or assist the doctor's diagnosis.
- Hypothesis: we will train and test a type of hypergraph CNN to enhance the performance of this task. And visualize the result of this stage output to testify.