Khalifa University of Science and Technology, PO Box 127788, Abu Dhabi, United Arab Emirates.
University of Louisville, Louisville, KY, 40292, USA.
J Digit Imaging. 2019 Oct;32(5):793-807. doi: 10.1007/s10278-018-0160-1.
We address the problem of prostate lesion detection, localization, and segmentation in T2W magnetic resonance (MR) images. We train a deep convolutional encoder-decoder architecture to simultaneously segment the prostate, its anatomical structure, and the malignant lesions. To incorporate the 3D contextual spatial information provided by the MRI series, we propose a novel 3D sliding window approach, which preserves the 2D domain complexity while exploiting 3D information. Experiments on data from 19 patients provided for the public by the Initiative for Collaborative Computer Vision Benchmarking (I2CVB) show that our approach outperforms traditional pattern recognition and machine learning approaches by a significant margin. Particularly, for the task of cancer detection and localization, the system achieves an average AUC of 0.995, an accuracy of 0.894, and a recall of 0.928. The proposed mono-modal deep learning-based system performs comparably to other multi-modal MR-based systems. It could improve the performance of a radiologist in prostate cancer diagnosis and treatment planning.
我们解决了 T2W 磁共振（MR）图像中前列腺病变检测、定位和分割的问题。我们训练了一个深度卷积编解码器架构，以同时分割前列腺、其解剖结构和恶性病变。为了结合 MRI 序列提供的 3D 上下文空间信息，我们提出了一种新颖的 3D 滑动窗口方法，该方法在利用 3D 信息的同时保留了 2D 域的复杂性。在 Initiative for Collaborative Computer Vision Benchmarking (I2CVB) 提供的 19 名患者的数据上进行的实验表明，我们的方法明显优于传统的模式识别和机器学习方法。特别是，对于癌症检测和定位任务，该系统的平均 AUC 为 0.995，准确性为 0.894，召回率为 0.928。所提出的基于单模态深度学习的系统与其他基于多模态 MR 的系统性能相当。它可以提高放射科医生在前列腺癌诊断和治疗计划中的表现。
