关键词: 弥漫大B细胞淋巴瘤, 生发中心B细胞样, 非GCB, 影像组学, 深度学习

Abstract:

Background and purpose: Diffuse large B-cell lymphoma (DLBCL) is subclassified into germinal center B-cell-like (GCB) and non-GCB subtypes, which differ in prognosis and treatment response. However, current distinction still relies on invasive pathological assays. This study developed radiomics and deep-learning models based on multiparametric magnetic resonance imaging (MRI) to non-invasively differentiate the two subtypes preoperatively, thereby reducing dependence on histopathological examination. Methods: This study retrospectively included patients with pathologically confirmed DLBCL diagnosed at Huashan Hospital, Fudan University, and other institutions between March 2013 and December 2024. Using multiparametric MRI data, we developed DLBCL-subtype classification models that combined 4 radiomics-based machine-learning algorithms: support vector machine (SVM), logistic regression (LR), Gaussian process (GP) and Naive Bayes (NB), with 3 deep-learning architectures [densely-connected convolutional networks 121 (DenseNet121), residual network 101 (ResNet101) and EfficientNet-b5]. Additionally, two radiologists with different experience levels independently classified DLBCL on MRI in a blinded fashion. Model and radiologist performance were quantified using the area under the receiver operating characteristic curve (AUC), accuracy (ACC), and F1-score to evaluate their ability to distinguish GCB from non-GCB subtypes. This study was approved by the Ethics Committee of Huashan Hospital of Fudan University (No. KY2024-663), and all patients signed informed consents. Results: A total of 173 patients were enrolled (55 with GCB subtype and 118 with non-GCB subtype). Radiomics and deep learning methods effectively distinguished DLBCL subtypes. Among these, the GP radiomics model (based on T1-CE+T2-FLAIR+ADC sequences) and DenseNet121 deep learning model (based on T1-CE+T2-FLAIR+ADC sequences) demonstrated optimal performance. Both achieved excellent results on the internal validation set (GP: AUC=0.900, ACC=0.896, F1=0.840; DenseNet121: AUC=0.846, ACC=0.854, F1=0.774) and maintained robustness on the external validation set. Furthermore, the classification efficacy of the optimal AI model surpassed that of experienced radiologists (highest physician AUC=0.678). Conclusion: Radiomics and deep-learning models based on multiparametric MRI features can effectively differentiate GCB from non-GCB subtypes of DLBCL. Among them, GP and DenseNet121 exhibit outstanding performance, especially when integrating multi-sequence feature sets for classifying DLBCL subtypes on complex imaging data.

Key words: Diffuse large B-cell lymphoma, Germinal center B-cell-like, Non-GCB, Radiomics, Deep learning