ESTRO 2024 - Abstract Book

S5061

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

1539

Digital Poster

Deep-learning based cell identification to predict colorectal cancer patient prognosis

boshen zhang 1,2,3 , jiazhou wang 1,2,3 , zhen zhang 1,2,3

1 Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai 200032, China. 2 Department of Oncology, Shanghai Medical College, Fudan University, Shanghai 200032, China. 3 Shanghai Clinical Research Center for Radiation Oncology, Shanghai Key Laboratory of Radiation Oncology, Shanghai 200032, China

Purpose/Objective:

Cancer prognosis research is a field focused on predicting the disease progression and outcomes of cancer patients. Due to the continuous increase in cancer incidence and mortality rates in recent years, studying prognosis prediction has become a significant challenge. Accurate analysis of cancer prognosis holds great significance in the treatment of cancer patients and reducing cancer mortality. In this study, based on Whole Slide Images (WSI) of pathology slides, we developed a two-stage prognosis prediction model using deep learning and machine learning methods

Material/Methods:

The two-stage model proposed in this article consists of the following: Initially, a convolutional neural network is trained to identify fibroblasts, tumor cells, and tumor-infiltrating lymphocytes within Whole Slide Images (WSI) slices. This model is referred to as the WSI Cell Identification Model. Using this model, the mentioned cell types within WSI slices can be identified. Based on the identification results, a machine learning model for feature extraction and cell classification is developed. This model extracts physical features of the cells identified by the first model, clusters them based on these features, calculates the relative quantities of different types of fibroblasts in each WSI, and relates them to patient prognosis outcomes. In this study, only the physical features of fibroblasts were extracted. Three colorectal cancer databases were used in this research: TCGA-READ, TCGA-COAD, and data from Fudan University Tumor Hospital. Among these, there were 148 and 378 patients with a total of 1,278 pathology slides in READ and COAD, while Fudan University Tumor Hospital's data included 81 patients and 145 pathology slides. The training and validation of the cell identification model were conducted on 70% training data, 20% validation data, and 10% testing data in COAD and READ, with external validation on Fudan University Tumor Hospital's dataset. In the prognosis prediction process, single-factor analysis, multi-factor analysis, and survival analysis were performed. To make the results more clinically relevant and closer to real-world scenarios, an analysis was also conducted based on different stages of cancer, aligning with clinical stage.

Results:

In terms of cell identification, this article achieved an accuracy of 0.78 for fibroblast identification, 0.88 for tumor infiltrating lymphocyte identification, and 0.79 for tumor cell identification. Simultaneously, significant results related to patient prognosis were demonstrated. In the COAD dataset, the p-values for survival analysis from stage I to IV and