Huang Huan, Qiu Liheng, Yang Shenmiao, Li Longxi, Nan Jiaofen, Li Yanting, Han Chuang, Zhu Fubao, Zhao Chen, Zhou Weihua
School of Computer Science and Technology, Zhengzhou University of Light Industry, Zhengzhou, Henan, China.
Department of Nuclear Medicine, Peking University People's Hospital, Beijing, China.
Med Phys. 2025 Mar 20. doi: 10.1002/mp.17763.
Accurate segmentation of diffuse large B-cell lymphoma (DLBCL) lesions is challenging due to their complex patterns in medical imaging. Traditional methods often struggle to delineate these lesions accurately.
This study aims to develop a precise segmentation method for DLBCL using 18F-fluorodeoxyglucose (18F-FDG) positron emission tomography (PET) and computed tomography (CT) images.
We propose a 3D segmentation method based on an encoder-decoder architecture. The encoder incorporates a dual-branch design based on the shifted window transformer to extract features from both PET and CT modalities. To enhance feature integration, we introduce a multi-scale information fusion (MSIF) module that performs multi-scale feature fusion using cross-attention mechanisms with a shifted window framework. A gated neural network within the MSIF module dynamically adjusts feature weights to balance the contributions from each modality. The model is optimized using the dice similarity coefficient (DSC) loss function, minimizing discrepancies between the model prediction and ground truth. Additionally, we computed the total metabolic tumor volume (TMTV) and performed statistical analyses on the results.
The model was trained and validated on a private dataset of 165 DLBCL patients and a publicly available dataset (autoPET) containing 145 PET/CT scans of lymphoma patients. Both datasets were analyzed using five-fold cross-validation. On the private dataset, our model achieved a DSC of 0.7512, sensitivity of 0.7548, precision of 0.7611, an average surface distance (ASD) of 3.61 mm, and a Hausdorff distance at the 95th percentile (HD95) of 15.25 mm. On the autoPET dataset, the model achieved a DSC of 0.7441, sensitivity of 0.7573, precision of 0.7427, ASD of 5.83 mm, and HD95 of 21.27 mm, outperforming state-of-the-art methods (p < 0.05, t-test). For TMTV quantification, Pearson correlation coefficients of 0.91 (private dataset) and 0.86 (autoPET) were observed, with R values of 0.89 and 0.75, respectively. Extensive ablation studies demonstrated the MSIF module's contribution to enhanced segmentation accuracy.
This study presents an effective automatic segmentation method for DLBCL that leverages the complementary strengths of PET and CT imaging. The method demonstrates robust performance on both private and publicly available datasets, ensuring its reliability and generalizability. Our method provides clinicians with more precise tumor delineation, which can improve the accuracy of diagnostic interpretations and assist in treatment planning for DLBCL patients. The code for the proposed method is available at https://github.com/chenzhao2023/lymphoma_seg.
弥漫性大B细胞淋巴瘤（DLBCL）病变在医学影像中具有复杂的模式，其准确分割具有挑战性。传统方法往往难以准确勾勒出这些病变。
本研究旨在利用18F-氟脱氧葡萄糖（18F-FDG）正电子发射断层扫描（PET）和计算机断层扫描（CT）图像开发一种用于DLBCL的精确分割方法。
我们提出了一种基于编码器-解码器架构的3D分割方法。编码器采用基于移位窗口变换器的双分支设计，以从PET和CT模态中提取特征。为了增强特征融合，我们引入了一个多尺度信息融合（MSIF）模块，该模块使用具有移位窗口框架的交叉注意力机制执行多尺度特征融合。MSIF模块内的门控神经网络动态调整特征权重，以平衡各模态的贡献。该模型使用骰子相似系数（DSC）损失函数进行优化，最小化模型预测与地面真值之间的差异。此外，我们计算了总代谢肿瘤体积（TMTV）并对结果进行了统计分析。
该模型在一个包含165例DLBCL患者的私有数据集和一个包含145例淋巴瘤患者PET/CT扫描的公开可用数据集（autoPET）上进行了训练和验证。两个数据集均使用五折交叉验证进行分析。在私有数据集上，我们的模型实现了DSC为0.7512，灵敏度为0.7548，精度为0.7611，平均表面距离（ASD）为3.61mm，第95百分位数的豪斯多夫距离（HD95）为15.25mm。在autoPET数据集上，该模型实现了DSC为0.7441，灵敏度为0.7573，精度为0.7427，ASD为5.
