Zabihollahy Fatemeh, Wu Holden H, Sisk Anthony E, Reiter Robert E, Raman Steven S, Fleshner Neil E, Yousef George M, Sung KyungHyun
Department of Radiological Sciences, David Geffen School of Medicine at UCLA, Los Angeles, Calif.
Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, Canada.
Radiol Imaging Cancer. 2025 May;7(3):e240336. doi: 10.1148/rycan.240336.
Purpose To develop and evaluate a novel deep learning-based approach for registering presurgical MR and whole-mount histopathology (WMHP) images of the prostate. Materials and Methods This retrospective study included patients who underwent prostate MRI before radical prostatectomy between July 2016 and June 2020. High-resolution ex vivo MRI was used as a reference to assess the structural relationship between in vivo MRI and WMHP. An Anatomy-Aware Morph model, a hybrid attention and convolutional neural network-based approach, was developed for multimodality prostate image registration. The pipeline included a module to estimate and correct distortion and motion between the prostate specimen and outside the human body. The dataset was divided into 270 and 45 patients for training and testing, respectively. Registration accuracy was evaluated using Dice similarity coefficient (DSC), Hausdorff distance, and target registration error. Results The proposed approach was validated using 160 images extracted from 45 male patients in the testing dataset with the average age ± SD of 64.0 years ± 6.6. The method achieved a DSC and Hausdorff distance of 0.95 ± 0.06 and 1.84 mm ± 0.38. The two-dimensional target registration errors between 90 sets of landmarks on in vivo MR images and WMHP images were 3.93 mm ± 0.80 and 1.18 mm ± 0.28 before and after registration ( < .001). The developed algorithm significantly outperformed the state-of-the-art VoxelMorph method for multimodality prostate image registration ( < .0001 for both DSC and Hausdorff distance). Conclusion The developed registration method successfully aligned presurgical prostate MR and histopathology images, facilitating automated mapping of prostate cancer from WMHP to MRI. Affine Transformation, Deformable Registration, Prostate Magnetic Resonance Imaging, Prostate Whole-Mount Histopathology © RSNA, 2025.
目的 开发并评估一种基于深度学习的新型方法，用于配准前列腺手术前的磁共振成像（MR）和全层组织病理学（WMHP）图像。材料与方法 这项回顾性研究纳入了2016年7月至2020年6月期间接受根治性前列腺切除术前行前列腺MRI检查的患者。高分辨率离体MRI用作参考，以评估活体MRI与WMHP之间的结构关系。开发了一种基于注意力和卷积神经网络的混合解剖感知形态模型，用于多模态前列腺图像配准。该流程包括一个模块，用于估计和校正前列腺标本与人体外部之间的畸变和运动。数据集分别分为270例和45例患者用于训练和测试。使用骰子相似系数（DSC）、豪斯多夫距离和目标配准误差评估配准准确性。结果 在测试数据集中，从45名男性患者中提取的160幅图像对所提出的方法进行了验证，平均年龄±标准差为64.0岁±6.6岁。该方法的DSC和豪斯多夫距离分别为0.95±0.06和1.84 mm±0.38。活体MR图像和WMHP图像上90组地标点之间的二维目标配准误差在配准前后分别为3.93 mm±0.80和1.18 mm±0.28（<0.001）。所开发的算法在多模态前列腺图像配准方面明显优于当前最先进的VoxelMorph方法（DSC和豪斯多夫距离均<0.0001）。结论 所开发的配准方法成功地对齐了手术前前列腺MR和组织病理学图像，便于将前列腺癌从WMHP自动映射到MRI。仿射变换、可变形配准、前列腺磁共振成像、前列腺全层组织病理学 ©RSNA，2025年
