Mohammed R. S. Sunoqrot¹, Kirsten M. Selnæs^1,2, Elise Sandsmark², Sverre Langørgen², Helena Bertilsson^3,4, Tone F. Bathen^1,2, and Mattijs Elschot^1,2
1Department of Circulation and Medical Imaging, NTNU, Norwegian University of Science and Technolog, Trondheim, Norway, ²Department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway, ³Department of Cancer Research and Molecular, NTNU, Norwegian University of Science and Technolog, Trondheim, Norway, ⁴Department of Urology, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway

Organ segmentation is an essential step in computer-aided diagnosis systems. Deep learning (DL)-based methods provide good performance for prostate segmentation, but little is known about their repeatability. In this work, we investigated the intra-patient repeatability of shape features for DL-based segmentation methods of the whole prostate (WP), peripheral zone (PZ) and transition zone (TZ) on T2-weighted MRI, and compared it to the repeatability of manual segmentations. We found that the repeatability of the investigated methods is comparable to manual for most of the investigated shape features from the WP and TZ segmentations, but not for PZ segmentations in all methods.

Jeremiah Sanders¹, Henry Mok², Alexander Hanania³, Aradhana Venkatesan⁴, Chad Tang², Teresa Bruno², Howard Thames⁵, Rajat Kudchadker⁶, and Steven Frank^2
1Imaging Physics, UT MD Anderson Cancer Center, Houston, TX, United States, ²Radiation Oncology, UT MD Anderson Cancer Center, Houston, TX, United States, ³Radiation Oncology, Baylor College of Medicine, Houston, TX, United States, ⁴Diagnostic Radiology, UT MD Anderson Cancer Center, Houston, TX, United States, ⁵Biostatistics, UT MD Anderson Cancer Center, Houston, TX, United States, ⁶Radiation Physics, UT MD Anderson Cancer Center, Houston, TX, United States

Quantitative techniques for characterizing deep learning (DL) algorithms are necessary to inform their clinical application, use, and quality assurance. This work analyzes the performance of DL algorithms for segmentation in prostate MRI at a population level. We performed computational observer studies and spatial entropy mapping for characterizing the variability of DL segmentation algorithms and evaluated them on a clinical MRI task that informs the treatment and management of prostate cancer patients. Specifically, we analyzed the task of prostate and peri-prostatic anatomy segmentation in prostate MRI and compared human and computer observer populations against one another.

Han Yu¹, Varut Vardhanabhuti¹, and Peng Cao^1
1The University of Hong Kong, Hong Kong, Hong Kong

We propose a novel gradient-based meta-learning scheme to tackle the challenges when deploying the model to a different medical center with the lack of labeled data. A pre-trained model is always suboptimal when deploying to different medical centers, where various protocols and scanners are used. Our method combines a 2D U-Net as a segmentor to generate segmentation maps and an adversarial network to learn from the shape prior in the meta-train and meta-test. Evaluation results on the public prostate MRI data and our HKU local database show that our approach outperformed the existing naive U-Net methods.

Piqiang Li¹, Zhao Li², Qinjia Bao², Kewen Liu¹, Xiangyu Wang³, Guangyao Wu⁴, and Chaoyang Liu^2
1School of Information Engineering, Wuhan University of Technology, Wuhan, China, ²State Key Laboratory of Magnetic Resonance and Atomic and Molecular Physics, Wuhan Institute of Physics and Mathmatics, Innovation Academy for Precision Measurement Science and Technology, Wuhan, China, ³Department of Radiology, The First Affiliated Hospital of Shenzhen University, Shenzhen, China, ⁴Department of Radiology, Shenzhen University General Hospital, Shenzhen, China

We proposed a Mutual Communicated Deep learning Segmentation and Classification Network (MC-DSCN) for prostate cancer based on multi-parametric MRI. The network consists of three mutual bootstrapping components: the coarse segmentation component provides coarse-mask information for the classification component, the mask-guided classification component based on multi-parametric MRI generates the location maps, and the fine segmentation component guided by the located maps. By jointly performing segmentation based on pixel-level information and classification based on image-level information, both segmentation and classification accuracy are improved simultaneously.

Cheng Li¹, Hui Sun¹, Taohui Xiao¹, Xin Liu¹, Hairong Zheng¹, and Shanshan Wang^1
1Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China

Automatic prostate MR image segmentation is needed to help doctors achieve fast and accurate disease diagnosis and treatment planning. Deep learning (DL) has shown promising achievements. However, DL models often face challenges in applications when there are large discrepancies between the training (source domain) and test (target domain) data. Here we propose a novel unsupervised domain adaptation method to address this issue without utilizing any target domain labels. Our method introduces two models trained in parallel to filter and correct the pseudo-labels generated for the target domain training data and thus, achieves substantially improved segmentation results on the test data.

Nader Aldoj¹, Federico Biavati¹, Sebastian Stober², Marc Dewey¹, Patrick Asbach¹, and Ingolf Sack^1
1Charité, Berlin, Germany, ²Ovgu Magdeburg, Magdeburg, Germany

The purpose was to investigate the impact of individual or combined MR Elastogrphy maps and MRI sequences on the overall segmentation of prostate gland and its subsequent zones using dense-like U-net. Our study showed that the obtained dice score of MRE maps was higher (i.e. more accurate segmentation) than the one obtained with MRI sequences. Moreover, we found that the magnitude MRE map had the highest importance for accurate segmentation among all tested maps/sequences. In conclusion, MRE maps resulted in excellent segmentations even when compared to T2w images which are the standard choice for segmentation tasks. 

Harri Merisaari¹, Pekka Taimen², Otto Ettala², Juha Knaapila², Kari T Syvänen², Esa Kähkönen², Aida Steiner², Janne Verho², Paula Vainio², Marjo Seppänen³, Jarno Riikonen⁴, Sanna Mari Vimpeli⁴, Antti Rannikko⁵, Outi Oksanen⁵, Tuomas Mirtti⁵, Ileana Montoya Perez¹, Tapio Pahikkala¹, Parisa Movahedi¹, Tarja Lamminen², Jani Saunavaara², Peter J Boström², Hannu J Aronen¹, and Ivan Jambor^6
1University of Turku, Turku, Finland, ²Turku University Hospital, Turku, Finland, ³Satakunta Central Hospital, Pori, Finland, ⁴Tampere University Hospital, Tampere, Finland, ⁵Helsinki University Hospital, Helsinki, Finland, ⁶Icahn School of Medicine at Mount Sinai, New York, NY, United States

PRORAD is a series of machine learning challenges hosted at CodaLab which provide access to prostate MRI data sets from 4 centers performed using a publicly available IMPROD bpMRI acquisition protocol. The challenge is designed for purposes of developing, validating and independent testing of various machine learning methods for prostate MRI. 
 
 

Francesco Giganti¹, Eoin Dinneen¹, Veeru Kasivisvanathan¹, Aiman Haider², Alex Freeman², Mark Emberton¹, Greg Shaw², Caroline M Moore¹, and Clare Allen^2
1University College London, London, United Kingdom, ²University College London Hospital, London, United Kingdom

We assessed the interobserver reproducibility of the Prostate Imaging Quality (PI-QUAL) score for prostate MR quality between two expert radiologists who independently scored a total of 41 multiparametric prostate MR scans from different vendors and scanners. All men included in this study had biopsy-confirmed prostate cancer and received robotic-assisted laparoscopic prostatectomy after imaging. Agreement was substantial (κ = 0.77; percent agreement = 80%) when assessing each single PI-QUAL score (1 to 5). Two expert radiologists achieved substantial reproducibility for the PI-QUAL score, but the composition of the scoring system will need to undergo further refinements.

Guangyu Dan^1,2, Min Li³, Mingshuai Wang⁴, Zheng Zhong^1,2, Kaibao Sun¹, Muge Karaman^1,2, Tao Jiang³, and Xiaohong Joe Zhou^1,2,5
1Center for MR Research, University of Illinois at Chicago, Chicago, IL, United States, ²Department of Bioengineering, University of Illinois at Chicago, Chicago, IL, United States, ³Department of Radiology, Beijing Chaoyang Hospital, Capital Medical University, Beijing, China, ⁴Department of Urology, Beijing Chaoyang Hospital, Capital Medical University, Beijing, China, ⁵Departments of Radiology and Neurosurgery, University of Illinois at Chicago, Chicago, IL, United States

Diffusion-weighted signal attenuation pattern contains valuable information regarding diffusion properties of the underlying tissue microstructures. With their extraordinary pattern recognition capability, deep learning (DL) techniques have a great potential to analyze diffusion signal decay. In this study, we proposed a 3D residual convolutional neural network (R3D) to detect prostate cancer by embedding the diffusion signal decay into one of the convolutional dimensions. By combining R3D with multi-task learning (R3D_MT), an excellent and stable prostate cancer detection performance was achieved in the peripheral zone (AUC of 0.990±0.008) and the transitional zone (AUC of 0.983±0.016).

Dharmesh Singh¹, Virendra Kumar², Chandan J Das³, Anup Singh¹, and Amit Mehndiratta^1
1Centre for Biomedical Engineering (CBME), Indian Institute of Technology (IIT) Delhi, New Delhi, India, ²Department of NMR, All India Institute of Medical Sciences (AIIMS) Delhi, New Delhi, India, ³Department of Radiology, All India Institute of Medical Sciences (AIIMS) Delhi, New Delhi, India

Automatic grading of prostate cancer (PCa) can play a major role in its early diagnosis, which has a significant impact on patient survival rates. The objective of this study was to develop and validate a framework for classification of PCa grades using texture features of T2-weighted MR images. Evaluation of classification result shows accuracy of 85.10 ± 2.43% using random forest feature selection and Gaussian support-vector machine classifier.

guiqin LIU¹, Guangyu Wu¹, yongming Dai², Ke Xue², and Shu Liao^3
1Radiology, Renji Hospital，Shanghai Jiaotong University School of Medicine, Shanghai, China, ²United Imaing Healthcare, Shanghai, China, ³Shanghai United Imaging Intelligence Co. Ltd, Shanghai, China

The interpretation of mpMRI is limited by expertise required and interobserver variability. Here we present an AI model, with ordinary accuracy level for diagnosing prostate cancer, the remarkable of false negative and sensitivity could help to reduce missed-diagnosis of PCa especially csPCa. To assess its performance in clinical setting, we curated internal and external validation, and performance of radiologists, AI model, human-machine synergy were compared, although AI performed suboptimal, the human-led synergy method performed equivalent to clinical assessment with improved consistency, which can serve as a comparison standard to more complex deep learning and synergy approaches in the future.

Dyah Ekashanti Octorina Dewi¹, Mohammed R. S. Sunoqrot¹, Gabriel Addio Nketiah¹, Elise Sandsmark², Sverre Langørgen², Helena Bertilsson^3,4, Mattijs Elschot^1,2, and Tone Frost Bathen^1,2
1Department of Circulation and Medical Imaging, Norwegian University of Science and Technology, Trondheim, Norway, ²Department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway, ³Department of Cancer Research and Molecular Medicine, Norwegian University of Science and Technology, Trondheim, Norway, ⁴Department of Urology, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway

Identifying repeatability of radiomic features in T2-Weighted prostate MRI is important to develop consistent imaging biomarkers and evaluate prostate cancer. This study aims to investigate the impact of pre-processing configurations on radiomic features repeatability in two short-term measurements. Repeatability was analyzed through pre-processing combinations of quantization, bin number, normalization, and filtering, and PyRadiomics feature extraction tool. The results show that repeatability of texture features varies depending on anatomical zones and lesions, pre-processing, and feature itself. Four texture features provide good repeatability independently of pre-processing configuration. Moreover, different feature groups have the highest repeatability in whole prostate, prostate zones and lesions.

Masami Yoneyama¹, Takashige Yoshida², Jihun Kwon¹, Kohei Yuda², Yuki Furukawa², Nobuo Kawauchi², Johannes M Peeters³, and Marc Van Cauteren^3
1Philips Japan, Tokyo, Japan, ²Radiology, Tokyo Metropolitan Police Hospital, Tokyo, Japan, ³Philips Healthcare, Best, Netherlands

Compressed SENSE-AI, based on Adaptive-CS-Net, clearly reduces noise artifacts and significantly improves the accuracy and robustness of T2 values in submillimeter (0.7mm) high-resolution prostate multi-echo turbo spin-echo T2 mapping compared with conventional SENSE and Compressed SENSE techniques, without any penalty for scan parameters. This technique may prove of value in better discriminating prostate cancer from healthy tissues. 

Rakesh Shiradkar¹, Michael Sobota¹, Leonardo Kayat Bittencourt², Sreeharsha Tirumani², Justin Ream³, Ryan Ward³, Amogh Hiremath¹, Ansh Roge¹, Amr Mahran¹, Andrei Purysko³, Lee Ponsky², and Anant Madabhushi^1
1Case Western Reserve University, Cleveland, OH, United States, ²University Hospitals Cleveland Medical Center, Cleveland, OH, United States, ³Cleveland Clinic, Cleveland, OH, United States

Radiomic approaches for prostate cancer risk stratification largely depend on radiologist delineation of prostate cancer regions of interest (ROI) on MRI. In this study, we acquired multi-reader delineations of ROIs, derived radiomic features within the ROIs trained and evaluated machine learning classifiers. We observed that variation in delineations did not affect the classification performance within a cohort but it did affect when evaluated on an independent validation set. We observed that a more conservative approach in delineations may ensure better generalizability and classification performance of machine learning models.

Chao Han¹, Shuai Ma¹, Xiang Liu¹, Yi Liu¹, Changxin Li², Yaofeng Zhang², Xiaodong Zhang¹, and Xiaoying Wang^1
1Department of Radiology, Peking University First Hospital, Beijing, China, ²Beijing Smart Tree Medical Technology Co. Ltd., Beijing, China

MR-based radiomics has been showed the feasibility in predicting high-grade prostate cancer (PCa), but most of the volumes of interest (VOIs) were based on manual segmentation. We develop and test 4 radiomics models based on manual/automatic segmentation of prostate gland/PCa lesion from apparent diffusion coefficient (ADC) maps to predict high-grade (Gleason score, GS ≥4+3) PCa at radical prostatectomy. Radiomics models based on automatic segmentation may obtain roughly the same diagnostic efficacy as manual segmentation and preoperative biopsy, which suggests the possibility of a fully automatic workflow combining automated segmentation and radiomics analysis.

Vanya V Valindria¹, Saurabh Singh², Eleni Chiou¹, Thomy Mertzanidou¹, Baris Kanber¹, Shonit Punwani², Marco Palombo¹, and Eleftheria Panagiotaki^1
1Centre for Medical Image Computing, Department of Computer Science, University College London, London, United Kingdom, ²Centre for Medical Imaging, University College London, London, United Kingdom

This study proposes non-invasive Gleason Score (GS) classification for prostate cancer with VERDICT-MRI using convolutional neural networks (CNNs). We evaluate GS classification using parametric maps from the VERDICT prostate model with compensated relaxation. We classify lesions using two CNN architectures: DenseNet and SE-ResNet. Results show that VERDICT achieves high GS classification performance using SE-ResNet with all parametric maps as input. Also in comparison with published GS classification multi-parametric MRI studies, VERDICT maps achieve higher metrics.  

Matthew Gibbons¹, Jeffry P Simko^2,3, Peter R Carroll², and Susan Noworolski^1
1Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, ²Urology, University of California, San Francisco, San Francisco, CA, United States, ³Pathology, University of California, San Francisco, San Francisco, CA, United States

Multi-parametric MRI has proven itself to be a clinically useful tool to assess prostate cancer through a qualitative assessment of multiple parameters within the guidelines of PI-RADs. Our objective is to leverage quantitative data combinations to characterize progression risk utilizing an automated procedure. This study showed the feasibility of MRI generated cancer risk maps, created from a combination of pre-prostatectomy, multiparametric MR images (mpMRI), to detect prostate cancer lesions >0.1cc as validated with histopathology. The method also quantified the volume of cancer within the prostate. Method improvements were identified by determining root causes for over and underestimation of cancer volumes.

Patricia M Johnson¹, Angela Tong¹, Paul Smereka¹, Awani Donthireddy¹, Robert Petrocelli¹, Hersh Chandarana¹, and Florian Knoll^1
1Center for Advanced Imaging Innovation and Research (CAI2R), Department of Radiology, New York University School of Medicine, New york, NY, United States

Early diagnosis and treatment of prostate cancer (PCa) can be curative, but the blood test for PSA is limited in detecting clinically significant PCa. Current abbreviated MR imaging protocols do not sufficiently reduce scan time for practical routine screening. In this work we extend a variational network (VN) deep learning image reconstruction method for accelerated clinical prostate images. Our results show that VN reconstructions of accelerated T2W images have comparable image quality to the current clinical protocol and require ≤1 minute of acquisition time, which can enable rapid screening prostate MRI.

Seb Harrevelt¹, Lieke Wildenberg², Dennis Klomp², C.A.T. van den Berg², Josien Pluim³, and Alexander Raaijmakers^1
1TU Eindhoven, Utrecht, Netherlands, ²UMC Utrecht, Utrecht, Netherlands, ³TU Eindhoven, Rossum, Netherlands

Ultra high-field MR images suffer from severe image inhomogeneity and artefacts due to the B1 field. Deep learning is a potential solution to this problem but training is difficult because no perfectly homogeneous 7T images exist that could serve as a ground truth. In this work, artificial training data has been created using numerically simulated 7T B1 fields, perfectly homogeneous 1.5T images and a signal model to add typical 7T B1 inhomogeneity on top of 1.5T images. A Pix2Pix model has been trained and tested on out-of-domain data where it out-performs classic bias field reducing algorithms.

lei hu¹, Jungong Zhao¹, Caixia Fu², and Thomas Benkert^3
1Department of Diagnostic and Interventional Radiology, Shanghai Jiao Tong University Affiliated Sixt, 上海, China, ²MR Application Development, Siemens Shenzhen magnetic Resonance Ltd, shenzhen, China, ³MR Application Predevelopment, Siemens Healthcare, Erlangen, Gernmany, Erlangen, Germany

We developed a supervised learning framework based on GAN in order to synthesize apparent diffusion coefficient maps (s-ADC) using full-FOV DWI images; zoomed-FOV ADC (z-ADC) served as the reference. Synthesized ADC using DWI with b=1000 mm2/s (S-ADCb1000) has statistically significant lower RMSE and higher PSNR, SSIM, and FSIM than s-ADCb50 and s-ADCb1500 (All P < 0.001). Both z-ADC and s-ADCb1000 had better reproducibility regarding quantitative ADC values in all evaluated tissues and better performance in tumor detection and classification than full-FOV ADC (f-ADC). A deep learning framework based on GAN is a promising method to synthesize realistic z-ADC sets with good image quality and accuracy in prostate cancer detection.