School of Computer Science and Engineering, The Hebrew University of Jerusalem, Jerusalem, Israel.
Department of Radiology, Hadassah Hebrew University Medical Center, Jerusalem, Israel.
Int J Comput Assist Radiol Surg. 2024 Mar;19(3):423-432. doi: 10.1007/s11548-023-03020-y. Epub 2023 Oct 5.
Radiological detection and follow-up of pancreatic cysts in multisequence MRI studies are required to assess the likelihood of their malignancy and to determine their treatment. The evaluation requires expertise and has not been automated. This paper presents MC3DU-Net, a novel multisequence cascaded pipeline for the detection and segmentation of pancreatic cysts in MRI studies consisting of coronal MRCP and axial TSE MRI sequences.
MC3DU-Net leverages the information in both sequences by computing a pancreas Region of Interest (ROI) segmentation in the TSE MRI scan, transferring it to MRCP scan, and then detecting and segmenting the cysts in the ROI of the MRCP scan. Both the voxel-level ROI of the pancreas and the segmentation of the cysts are performed with 3D U-Nets trained with Hard Negative Patch Mining, a new technique for class imbalance correction and for the reduction in false positives.
MC3DU-Net was evaluated on a dataset of 158 MRI patient studies with a training/validation/testing split of 118/17/23. Ground truth segmentations of a total of 840 cysts were manually obtained by expert clinicians. MC3DU-Net achieves a mean recall of 0.80 ± 0.19, a mean precision of 0.75 ± 0.26, a mean Dice score of 0.80 ± 0.19 and a mean ASSD of 0.60 ± 0.53 for pancreatic cysts of diameter > 5 mm, which is the clinically relevant endpoint.
MC3DU-Net is the first fully automatic method for detection and segmentation of pancreatic cysts in MRI. Automatic detection and segmentation of pancreatic cysts in MRI can be performed accurately and reliably. It may provide a method for precise disease evaluation and may serve as a second expert reader.
在多序列 MRI 研究中，需要进行放射学检测和随访胰腺囊肿，以评估其恶性可能性，并确定治疗方案。这种评估需要专业知识，并且尚未实现自动化。本文提出了 MC3DU-Net，这是一种新颖的多序列级联管道，用于检测和分割 MRI 研究中的胰腺囊肿，包括冠状位 MRCP 和轴位 TSE MRI 序列。
MC3DU-Net 利用 TSE MRI 扫描中胰腺的感兴趣区域（ROI）分割信息，将其传输到 MRCP 扫描中，然后在 MRCP 扫描的 ROI 中检测和分割囊肿。胰腺的体素级 ROI 和囊肿的分割都是使用经过 Hard Negative Patch Mining 训练的 3D U-Net 进行的，这是一种用于校正类不平衡和减少假阳性的新技术。
MC3DU-Net 在一个包含 158 例 MRI 患者研究的数据集上进行了评估，训练/验证/测试集的划分比例为 118/17/23。总共由专家临床医生手动获得了 840 个囊肿的总体分割。MC3DU-Net 在直径大于 5 毫米的胰腺囊肿上的平均召回率为 0.80±0.19，平均精度为 0.75±0.26，平均 Dice 评分为 0.80±0.19，平均 ASSD 为 0.60±0.53，这是临床相关的终点。
MC3DU-Net 是首个用于 MRI 中胰腺囊肿检测和分割的全自动方法。MRI 中胰腺囊肿的自动检测和分割可以准确可靠地完成。它可以为精确的疾病评估提供一种方法，并可以作为第二个专家读者。
