Development and validation of the ultrasound-based radiomics and deep learning prognostic models for diffuse large B-cell lymphoma.

Authors

Abstract

Lymphoma includes fatal hematological malignancies, with diffuse large B-cell lymphoma (DLBCL) as the most common aggressive non-Hodgkin lymphoma subtype. Accurate early identification of high-risk DLBCL patients is key for individualized treatment and improved outcomes. This study developed and validated two ultrasound-based radiomics and deep learning prognostic models for DLBCL, and evaluated their utility in risk stratification and survival prediction. A retrospective cohort of 149 pathologically confirmed DLBCL patients was randomly split into training (<i>n</i> = 75) and independent test (<i>n</i> = 74) groups. A radiomics model and a ResNet50-based deep learning model were built on the training group, with external validation in the test group. Kaplan–Meier curves compared overall survival (OS) between risk groups. Model performance was evaluated through AUC (the primary indicator), C-index (the discrimination ability), DCA (net clinical benefit), and calibration curves (prediction-observation agreement). Both of these models were able to effectively classify patients into different risk groups (<i>P</i> < 0.05). The AUC values of the radiomics model in the training group were 0.819 (1 year), 0.803 (3 years), and 0.821 (5 years); in the test group, they were 0.686, 0.747, and 0.756. The AUC values of the deep learning model in the training group were 0.771, 0.782, and 0.815; in the test group, it performed better in both short-term (1 year AUC: 0.840) and long-term (3 years: 0.832; 5 years: 0.764) periods. The C index indicated that both models had high heterogeneity. DCA showed that in the training group, the imaging-based model had a higher overall net benefit, while in the test group, the deep learning model performed better. The calibration curve indicated that the radiomics model had better prediction and calibration capabilities, and might be more reliable in evaluating long-term prognosis. The ultrasound-based radiomics and ResNet50 deep learning models developed in this study demonstrate robust efficacy in predicting OS and stratifying risk in DLBCL patients. Both models are non-invasive, objective, and clinically applicable, providing valuable complementary tools for individualized prognosis evaluation and aiding clinicians in making data-driven treatment decisions for DLBCL patients.

Tags

Topics