DLBCL Prognosis Prediction with Segment Anything Model and Intraclass Correlation Coefficients Consistency Analysis on PET images

Abstract:
Diffuse large B-cell lymphoma (DLBCL) is one of the most common non-Hodgkin lymphoma, and it has a high mortality rate because of its unpredictable prognosis. Many methods have been proposed to predict the DLBCL prognosis accurately. This study aims to validate that deep learning-based methods can be used for the DLBCL segmentation and PFS or OS prediction through the intraclass correlation coefficients (ICC) consistency analysis. In this study, the fine-tuned segment anything model (SAM) is used for the DLBCL segmentation in PET images, and pre-trained deep learning models are directly used to extract features in the PFS and OS predictions. We collect three datasets from three different imaging centers. One dataset is used for the SAM fine-tuning and prognosis prediction models training, and the other two are used for the validation. With our proposed framework, we can obtain the best prognosis prediction results on the testing datasets. The results indicate that our framework has potential in DLBCL prognosis prediction.