METHODS: Eight nnU-Net-based segmentation models with 2D or 3D configuration and various loss functions were trained on positive-only or positive-and-negative datasets, comprising axial and coronal T2-weighted MR images from 254 scans on 146 ADPKD patients with pancreatic cysts labeled independently by two radiologists. Model performance was evaluated on test subjects unseen in training, comprising 40 internal, 40 external, and 23 test-retest reproducibility ADPKD patients.
RESULTS: Two radiologists agreed on 52% of cysts labeled on training data, and 33%/25% on internal/external test datasets. The 2D model with a loss of combined dice similarity coefficient and cross-entropy trained with the dataset with both positive and negative cases produced an optimal dice score of 0.7 ± 0.5/0.8 ± 0.4 at the voxel level on internal/external validation and was thus used as the best-performing model. In the test-retest, the optimal model showed superior reproducibility (83% agreement between scan A and B) in segmenting pancreatic cysts compared to six expert observers (77% agreement). In the internal/external validation, the optimal model showed high specificity of 94%/100% but limited sensitivity of 20%/24%.
CONCLUSIONS: Labeling pancreatic cysts on T2 images of the abdomen in patients with ADPKD is challenging, deep learning can help the automated detection of pancreatic cysts, and further image quality improvement is warranted.
方法:在仅正负或正负数据集上训练了八个具有2D或3D配置和各种损失函数的基于nnU-Net的分割模型,包括来自146例ADPKD患者的254次扫描的轴向和冠状T2加权MR图像,这些患者由两名放射科医生独立标记。在训练中看不见的测试对象上评估模型性能,包括40个内部,40个外部,23例复检重复性ADPKD患者。
结果:两位放射科医师对训练数据上标记的囊肿有52%达成一致,以及内部/外部测试数据集上的33%/25%。具有组合骰子相似性系数和交叉熵的损失的2D模型用具有正和负两种情况的数据集训练,在内部/外部验证的体素水平上产生0.7±0.5/0.8±0.4的最佳骰子得分,因此被用作表现最好的模型。在重测中,与6名专家观察者(77%的一致性)相比,最佳模型在胰腺囊肿分割方面显示出较好的可重复性(扫描A和扫描B的一致性为83%).在内部/外部验证中,最佳模型的特异性高,为94%/100%,但灵敏度有限,为20%/24%.
结论:在ADPKD患者的腹部T2图像上标记胰腺囊肿具有挑战性,深度学习可以帮助自动检测胰腺囊肿,和进一步的图像质量改进是必要的。