Huan Minh Luu¹, Gyu-sang Yoo², Dong-Hyun Kim¹, Won Park², and Sung-Hong Park^1
1Magnetic Resonance Imaging Laboratory, Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Korea, Republic of, ²Department of Radiation Oncology, Samsung Medical Center, Seoul, Korea, Republic of

MR-only radiotherapy planning can reduce the radiation exposure from repeated CT scanning. Most researches focus on generating synthetic CT images from MR, but not the contouring of organs-of-interest in said images, which requires manual labor and expertise. In this study, we proposed CycleSeg, a CycleGAN-based network that can accomplish both tasks to streamline the process and reduce human efforts. Experiments showed that the proposed CycleSeg can generate realistic synthetic CT images along with accurate organ segmentation in the pelvis of patients with prostate cancer.

Yang Zhang^1,2, Weikang Li³, Zhao Zhang³, Yingnan Xue³, Yan-Lin Liu², Peter Chang², Daniel Chow², Ke Nie¹, Min-Ying Su², and Qiong Ye^3,4
1Department of Radiation Oncology, Rutgers-Cancer Institute of New Jersey, Robert Wood Johnson Medical School, New Brunswick, NJ, United States, ²Department of Radiological Sciences, University of California, Irvine, CA, United States, ³Department of Radiology, The First Affiliate Hospital of Wenzhou Medical University, Wenzhou, China, ⁴High Magnetic Field Laboratory, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, China

A bi-directional Convolutional Long Short Term Memory (CLSTM) Network was previously shown capable of differentiating prostate cancer and benign prostate hyperplasia (BPH) based DCE-MRI that acquired 40 time frame images. The purpose of this work was to investigate the diagnostic value of peritumoral tissues. Several different methods were used to expand peritumoral tissues surrounding the lesion, and they were used as the input to the diagnostic network. A total of 135 cases were analyzed, including 73 prostate cancer and 62 BPH. Based on 4-fold cross-validation, the region growing based ROI had the best performance, with a mean AUC of 0.89.

Yihong Zhang¹, Ying Hou², Jie Bao³, Yang Song¹, Yu-dong Zhang², Xu Yan⁴, and Guang Yang^1
1East China Normal University, Shanghai, China, ²the First Affiliated Hospital with Nanjing Medical University, Nanjing, China, ³the First Affiliated Hospital with Soochow University, Soochow, China, ⁴Siemens Healthcare, Shanghai, China

We proposed an algorithm to incorporate radiologist’s prior-knowledge about location of extension into a CNN model to diagnose the extracapsular extension of the prostate cancer from multiparametric MRI (mpMRI). The model was trained on 596 cases with ensemble learning before validated with an independent validation cohort of 150 cases and an external cohort of 103 cases. Our proposed model achieved an area under receiver operating characteristic curve (AUC) of 0.807/0.728 on the internal/external test cohort, which is better than the traditional model (AUC=0.746/0.723) and the clinical reports by two radiologists (AUC=0.725, 0.632/0.694, 0.712).

Alexandros Patsanis¹, Mohammed R. S. Sunoqrot ¹, Elise Sandsmark ², Sverre Langørgen ², Helena Bertilsson ^3,4, Kirsten M. Selnæs ^1,2, Hao Wang⁵, Tone F. Bathen ^1,2, and Mattijs Elschot ^1,2
1Department of Circulation and Medical Imaging, Norwegian University of Science and Technology - NTNU, Trondheim, Norway, ²Department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway, ³Department of Clinical and Molecular Medicine, Norwegian University of Science and Technology - NTNU, Trondheim, Norway, ⁴Department of Urology, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway, ⁵Department of Computer Science, Norwegian University of Science and Technology - NTNU, Gjøvik, Norway

Generative Adversarial Networks (GANs) were evaluated for detection and visualization of prostate cancer, proposing an automated end-to-end pipeline. Two GANs were trained and tested with T2-weighted images from an in-house dataset of 646 patients. The weakly-supervised GAN performed better (AUC=0.785) than unsupervised GAN (AUC=0.462). The performance of the GANs was dependent on pre-processing parameters. The PROSTATEx dataset (N=204) was used for external validation, giving an AUC of 0.642. The weakly-supervised GAN showed promise for detecting and localizing prostate cancer on T2W MRI, but further research is necessary to improve model performance and generalizability.

Yongkai Liu^1,2, Haoxin Zheng¹, Zhengrong Liang³, Miao Qi¹, Wayne Brisbane⁴, Leonard Marks⁴, Steven Raman¹, Robert Reiter⁴, Guang Yang⁵, and Kyunghyun Sung^1
1Department of Radiological Sciences, University of California, Los Angeles, Los Angeles, CA, United States, ²Physics and Biology in Medicine IDP, University of California, Los Angeles, Los Angeles, CA, United States, ³Departments of Radiology and Biomedical Engineering, Stony Brook University, Stony Brook, New York, NY, United States, ⁴Department of Urology, University of California, Los Angeles, Los Angeles, CA, United States, ⁵National Heart and Lung Institute, Imperial College London, London, United Kingdom

Accurate classification of prostate cancer (PCa) enables better prognosis and selection of treatment plans. We presented a textured-based deep learning method to enhance prostate cancer classification performance by enriching deep learning with prostate cancer texture information.

Ruiqi Yu¹, Ying Hou², Yang Song¹, Yu-dong Zhang², and Guang Yang^1
1Shanghai Key Laboratory of Magnetic Resonance, East China Normal University, Shanghai, China, ²Department of Radiology, the First Affiliated Hospital with Nanjing Medical University, Jiangsu, China

Prostate cancer is one of the most important causes of cancer-incurred deaths among males. The prostate imaging reporting and data system (PI-RADS) v2 standardizes the acquisition of multi-parametric magnetic resonance images (mp-MRI) and identification of clinically significant prostate cancer. We purposed a convolutional neural network which integrated an unsure data model (UDM) to predict the PI-RADS v2 score from mp-MRI. The model achieved an F1 score of 0.640, which is higher than that of the ResNet-50. On an independent test cohort of 146 cases, our model achieved an accuracy of 64.4%.

Deepa Darshini Gunashekar¹, Lars Bielak^1,2, Arnie Berlin³, Leonard Hägele¹, Benedict Oerther⁴, Matthias Benndorf⁴, Anca Grosu^2,4, Constantinos Zamboglou^2,4, and Michael Bock^1,2
1Dept.of Radiology, Medical Physics, Medical Center University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany, ²German Cancer Consortium (DKTK), Partner Site Freiburg, Freiburg, Germany, ³The MathWorks, Inc., Novi, MI, United States, ⁴Dept.of Radiology, Medical Center University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany

An explainable deep learning model was implemented to interpret the predictions of a convolution neural network (CNN) for prostate tumor segmentation. The CNN automatically segments the prostate gland and prostate tumors in multi-parametric MRI data using co-registered whole mount histopathology images as ground truth. For the interpretation of the CNN, saliency maps are generated by generalizing the Gradient Weighted Class Activation Maps method for prostate tumor segmentation. Evaluations on the saliency method indicate that the CNN was able to correctly localize the tumor and the prostate by targeting the pixels in the image deemed important for the CNN's prediction. 

Eleni Chiou^1,2, Francesco Giganti^3,4, Shonit Punwani⁵, Iasonas Kokkinos², and Eleftheria Panagiotaki^1,2
1Centre of Medical Image Computing, University College London, London, United Kingdom, ²Department of Computer Science, University College London, London, United Kingdom, ³Department of Radiology, UCLH NHS Foundation Trust, University College London, London, United Kingdom, ⁴Division of Surgery & Interventional Science, University College London, London, United Kingdom, ⁵Centre for Medical Imaging, Division of Medicine, University College London, London, United Kingdom

In this work we utilize unsupervised domain adaptation for prostate lesion segmentation on VERDICT-MRI. Specifically, we use an image-to-image translation method to translate multiparametric-MRI data to the style of VERDICT-MRI. Given a successful translation we use the synthesized data to train a model for lesion segmentation on VERDICT-MRI. Our results show that this approach performs well on VERDICT-MRI despite the fact that it does not exploit any manual annotations. 

Zhaohuan Zhang¹, Sohrab Afshari Mirak¹, Melina Hosseiny¹, Afshin Azadikhah¹, Amirhossein Mohammadian Bajgiran¹, Alan Priester², Kyunghyun Sung¹, Anthony Sisk³, Robert Reiter², Steven Raman¹, Dieter Enzmann¹, and Holden Wu^1
1Department of Radiological Sciences, UCLA, Los Angeles, CA, United States, ²Department of Urology, UCLA, Los Angeles, CA, United States, ³Department of Pathology, UCLA, Los Angeles, CA, United States

Diffusion-relaxation correlation spectrum imaging (DR-CSI) can map prostate microstructure in both ex vivo and in vivo imaging. However, the translation of prostate DR-CSI to in vivo applications faces technical challenges regarding trade-offs between scan time and accuracy. The goal of this study was to develop a data-driven systematic framework to evaluate and select subsampled DR-CSI echo time and b-values encoding schemes that reduce scan time while maintaining accuracy of estimated prostate microstructure parameters.

Sean McTavish¹, Anh T. Van², Kilian Weiss³, Johannes M. Peeters⁴, Marcus R. Makowski², Rickmer F. Braren², and Dimitrios C. Karampinos^2
1Technical University of Munich, Munich, Germany, ²Department of Diagnostic and Interventional Radiology, Technical University of Munich, Munich, Germany, ³Philips Healthcare, Hamburg, Germany, ⁴Philips Healthcare, Best, Netherlands

In order to improve resolution without increasing geometric distortions, there is an ongoing interest in multi-shot DWI in the prostate due to the high clinical significance of prostate DWI for tumor staging and therapy monitoring. However, intershot phase variations require phase error estimation and correction to reconstruct the multi-shot DWI data. Since free-breathing scans are normally used in prostate DWI, respiratory motion can be a significant source of phase errors in the prostate. The present work aims to characterize and investigate the link between respiratory motion and phase errors in the prostate.