Zhongwei Chen¹, Yang Zhang², Jiejie Zhou¹, Youfan Zhao¹, Haiwei Miao¹, Huiru Liu¹, Shuxin Ye¹, Nina Xu¹, Meihao Wang¹, and Min-Ying Su^2
1Department of Radiology, The First Affiliated Hospital of Wenzhou Medical University, wenzhou, China, ²Department of Radiological Sciences, University of California, Irvine, CA, United States

A total of 200 patients receiving both DCE-MRI and mammography were analyzed, including 146 malignant and 56 benign lesions. For each lesion, 3D tumor mask was done using fuzzy-C-means clustering algorithm. Three DCE parametric maps were generated, and the radiomics features were extracted from these maps by PyRadiomics. Five models were built based on DCE-MRI, mammography, and the combination. The BI-RADS score was obtained from the radiology reports for comparison. The model built based on all MRI and mammography features yielded the highest accuracy of 89.6%, and had significantly better diagnostic performance than BI-RADS using threshold of 4A or 4B.

Luu-Ngoc Do¹, Chae Yeong Im², Jae Hyuk Park², So Yeon Ki³, Ilwoo Park^2,4,5, and Hyo Soon Lim^2,3
1Department of Radiology, Chonnam National University, Gwangju, Korea, Republic of, ²College of Medicine, Chonnam National University, Gwangju, Korea, Republic of, ³Department of Radiology, Chonnam National University Hwasun Hospital, Hwasun, Korea, Republic of, ⁴Department of Artificial Intelligence Convergence, Chonnam National University, Gwangju, Korea, Republic of, ⁵Department of Radiology, Chonnam National University Hospital, Gwangju, Korea, Republic of

This study aims to explore the effectiveness of deep learning algorithms for distinguishing pure (noninvasive) ductal carcinoma in situ (DCIS) from invasive disease for patients showing DCIS in core-needle biopsy using MRI. Preoperative axial dynamic contrast-enhanced MRI data from 352 patients were used to train, validate and test the two-step convolutional neural network (CNN) utilizing a recurrent model. Our model produced an accuracy of 69.4% and AUC of 0.721. The comparison between the proposed model and a 2D or 3D model suggests that the sequential information may provide an important support for occult invasive cancer in patients diagnosed with DCIS.

Xia Wu^1,2,3, Zhou Liu⁴, Meng Wang⁴, Zhe Ren^1,2,3, Ya Ren⁴, Jie Wen⁴, Qian Yang⁴, Xin Liu^1,2,3, Hairong Zheng^1,2,3, and Na Zhang^1,2,3
1Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Synopsis, ShenZhen, China, ²Key Laboratory for Magnetic Resonance and Multimodality Imaging of Guangdong Province, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, ShenZhen, China, ³CAS key laboratory of health informatics, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, ShenZhen, China, ⁴Department of Radiology, National Cancer Center/Cancer Hospital and Shenzhen Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, ShenZhen, China

We have achieved preoperative classification of breast cancer by combination of pharmacokinetic parameters and texture features using machine learning. Using the information available in each feature space, an appropriate feature fusion method using information from the two feature spaces can help the classification process and improve diagnosis accuracy. Among them, SVM and KNN have better performance.

Yan-Lin Liu¹, Yang Zhang^1,2, Jeon-Hor Chen^1,3, Siwa Chan⁴, Jiejie Zhou⁵, Meihao Wang⁵, and Min-Ying Su^1
1Department of Radiological Sciences, University of California, Irvine, CA, United States, ²Department of Radiation Oncology, Rutgers-Cancer Institute of New Jersey, Robert Wood Johnson Medical School, New Brunswick, NJ, United States, ³Department of Radiology, E-Da Hospital and I-Shou University, Kaohsiung, Taiwan, ⁴Department of Medical Imaging, Taichung Tzu-Chi Hospital, Taichung, Taiwan, ⁵Department of Radiology, The First Affiliate Hospital of Wenzhou Medical University, Wenzhou, China

Two datasets were used, 107 cases for training and 94 cases for testing. Patients were classified into three subtypes: TN, HER2+, and (HR+/HER2-). Three heuristic DCE parametric maps were generated from DCE-MRI. PyRadiomics was applied to extract features. Five machine learning algorithms were implemented to build models. The classification accuracy in training dataset was 84.3%, 77.2%, 75.5%, 74.3%, 69.1% for SVM, Decision tree, LDA, KNN, Naïve Bayes, respectively. In binary classification for TN vs. Non-TN, accuracy was 91.0% in training and 88.2% in testing datasets. For HER2+ vs. HER2-, accuracy was 90.4% in training and 86.2% in testing datasets.

Snekha Thakran¹, Rakesh Kumar Gupta², and Anup Singh^1,3
1Centre for Biomedical Engineering, Indian Institute of Technology Delhi, New Delhi, India, Delhi, India, ²Department of Radiology, Fortis Memorial Research Institute, Haryana, Gurgaon, India, Delhi, India, ³Department of Biomedical Engineering, All India Institute of Medical Science, New Delhi, India, Delhi, India

Multi-parametric MRI(mp-MRI) has shown promising outcomes with high sensitivity and accuracy in the characterization of breast tumor. Quantitative analysis of mp-MRI and texture features with machine learning approach have also shown potential in improving accuracy of breast tumor classification. The objective of this study was to differentiate low-grade vs. high-grade breast tumor using machine learning with optimized feature vector obtained from mp-MRI data. The study included mp-MRI data of 35 patients with breast cancer. The combination of support-vector-machine(SVM) with Wrapper method using Adaptive-Boosting(AdaBoost) technique resulted in high sensitivity(0.94±0.07), specificity(0.80±0.05), and accuracy(0.90±5.48) in classification of low-grade vs. high-grade tumors. 

Chen Chen¹, Fabao Gao¹, and Xiaoyue Zhou^2
1Department of Radiology, West China Hospital, Chengdu, China, ²MR Collaboration, Siemens Healthineers Ltd., Shanghai, China

Axillary lymph node dissection (ALND) is the gold standard for evaluating axillary lymph node metastasis (ALNM), but ALND may not confer a survival advantage. Therefore, reliable, noninvasive approaches for preoperative prediction of ALNM have been needed. The use of machine learning (ML) in predicting ALNM in breast cancer patients has been reported. We have conducted a large-sample-size assessment and a meta-analysis of published studies concerning the diagnostic performance of ML-based MRI in predicting ALNM in breast cancer patients.

Zijian Zhou¹, David E. Rauch¹, Jong Bum Son¹, Benjamin C. Musall¹, Nabil A. Elshafeey², Jason B. White³, Mark D. Pagel⁴, Stacy Moulder³, and Jingfei Ma^1
1Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, ²Breast Imaging, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, ³Breast Medical Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, ⁴Cancer Systems Imaging, The University of Texas MD Anderson Cancer Center, Houston, TX, United States

Prediction of response to neoadjuvant systemic therapy for triple-negative breast cancer is important for patient management. Here we constructed a deep learning convolutional and recursive neural network ensemble for early prediction of pathologic complete response utilizing pre-treatment DCE and DWI breast MRIs. Images from 135 patients were partitioned into training/validation/testing groups with the ratio of 80/20/35. For the testing group, the network achieved an accuracy of 69%, with the sensitivity of 75% and specificity of 63%. The area under the receiver operating characteristic curve was 0.68.

Yang Zhang^1,2, Yan-Lin Liu², Ke Nie¹, Jiejie Zhou³, Siwa Chan⁴, Vivian Youngjean Park⁵, Min Jung Kim⁵, Zhongwei Chen³, Jeon-Hor Chen^2,4, Meihao Wang³, and Min-Ying Su^2
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, ⁴Department of Medical Imaging, Taichung Tzu-Chi Hospital, Taichung, Taiwan, ⁵Department of Radiology and Research Institute of Radiological Science, Severance Hospital, Yonsei University College of Medicine, Seoul, Korea, Republic of

Mask R-CNN and ResNet50 were implemented to search the entire breast MRI to identify suspicious lesions, and then to further evaluate their likelihood of malignancy. The dataset included 103 malignant and 73 benign lesions in 153 patients. In detection phase using Mask R-CNN, 101 malignant, 48 benign, and 130 normal enhancing tissues were detected as suspicious. When putting them into ResNet50 for characterization, 99 cancers were correctly diagnosed as malignant, and only 16 benign lesions and 16 normal enhancing tissues remained as likely malignant. The true positive rate was 99/103=96%, and many detected false positives were dismissed during classification phase.

Pablo F. Damasceno^1,2, Tatiana Kelil^1,2, Rutwik Shah^1,2, Bruno Astuto Arouche Nunes^1,2, Jason Crane^1,2, and Sharmila Majumdar^1,2
1Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States, ²Center for Intelligent Imaging, University of California San Francisco, San Francisco, CA, United States

Given its potential importance as a biomarker of breast cancer risk, a reliable and objective quantitative measurement of Fibroglandular tissue (FGT) with limited intra or inter-rater variability will be invaluable in clinical practice. Currently, the amount of FGT in breast MRIs is reported via a 4-level qualitative system. We investigate the relationship between these classes and the amount of FGT, obtained via deep-learning segmentations. We find that the distribution of FGT in these classes deviates significantly from quartiles, but more uniform distributions can be achieved by emulating the radiologist’s workflow during clinical reporting.

Lauren K Fang¹, Ana E Rodriguez-Soto¹, Summer J Batasin², Kathryn E Keenan³, and Rebecca A Rakow-Penner^2
1Radiology, University of California San Diego, La Jolla, CA, United States, ²University of California San Diego, La Jolla, CA, United States, ³National Institute of Standards and Technology, Boulder, CO, United States

Diffusion-weighted imaging (DWI) has shown potential to diagnose and monitor breast cancer, but is limited in part due to geometric distortion artifacts that affect echo-planar imaging. Generating distortion-free images is a key step towards its clinical implementation. Here, we investigated the effectiveness of artifact-reduction data collection strategies (i.e. parallel imaging DWI and reduced field-of-view DWI) in combination with reverse polarity gradient post-processing in correcting for geometric distortion artifacts.

Kay van der Hoogt¹, Robert-Jan Schipper¹, Ronni Wessels¹, Cees de Graaf¹, Arjan te Boekhorst¹, Leon ter Beek², Regina Beets-Tan¹, and Ritse Mann^1
1Radiology, the Netherlands Cancer Institute – Antoni van Leeuwenhoek, Amsterdam, Netherlands, ²Medical Physics, the Netherlands Cancer Institute – Antoni van Leeuwenhoek, Amsterdam, Netherlands

It is advocated to perform breast DWI-acquisitions before contrast administration [1]. However, performing DWI after contrast MRI might be more practical in a clinical context. Several studies [2] evaluated whether performing breast MRI before or after contrast administration impacts the DWI classifying properties. However, previous research was often based on an interpatient analysis. Therefore, we performed a retrospective intra-patient analysis on perfusion-free ADC-maps before and after contrast administration.   A decrease in ADC was observed in post-contrast ADC-values. This was independent of the B₀-field  strength, and similar for benign and malignant lesions.

 

Muzhen He¹, Huiping Ruan¹, Mingping Ma¹, Zhongshuai Zhang², and Robert Grimm^2
1Radiology, Provincial Clinical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, China, ²Siemens Healthcare Ltd, Shanghai, China

This study aims to analyze the value of ADC, IVIM and DKI in differentiating benign and malignant breast nodules. The combination of all the quantitative diffusion parameters is also evaluated. The value of ADC, IVIM-D, KDI-K and DKI-D can be used as a diagnostic tool for evaluating benign and malignant breast lesions. Compared with IVIM-D, DKI-K and DKI-D, ADC value is the best single discriminative parameters to differentiate breast lesions. Further, the combination of ADC and DKI-K value shows the best diagnostic efficacy.

Mami Iima^1,2, Aika Okazawa³, Ryosuke Okumura³, Sachiko Takahara⁴, Tomotaka Noda³, Taro Nishi³, Yuji Nakamoto¹, and Masako Kataoka^1
1Diagnostic Imaging and Nuclear Medicine, Kyoto University Graduate School of Medicine, Kyoto, Japan, ²Clinical Innovative Medicine, Institute for Advancement of Clinical and Translational Science, Kyoto University Hospital, Kyoto, Japan, ³Radiology, Kitano Hospital, The Tazuke Kofukai Medical Research Institute, Osaka, Japan, ⁴Breast Surgery, Kitano Hospital, The Tazuke Kofukai Medical Research Institute, Osaka, Japan

Our proposed DWI reading method based on the BI-RADS lexicons from multiple b-value images had comparable specificity and NPV to the standard BI-RADS. The DWI reading method might increase diagnostic confidence in differentiating malignant and benign breast tumors. Excellent to substantial agreement was observed for DWI reading.  Substantial-perfect agreement was observed in lesion characteristics and fibrograndular tissues, according to an adjusted BI-RADS lexicon for lesion classification. These results suggest that DWI reading methods might be applicable in a clinical setting, however, the agreement is moderate for non-mass lesions, which is still challenging for clinical application.

Haiyun Wang¹, Qian Xu¹, Gaofeng Shi¹, Lijia Wang¹, Qinglei Shi², and Chen Zhang^2
1CTMRI, The Fourth Hospital of Hebei Medical University, Shijiazhuang,Hebei, China, ²MR Scientific Marketing,Siemens Healthcar, Beijing, China

In this study, we studied the feasibility of GRASP based DCE technique and readout-segmented DWI technique in early prediction and evaluation of breast cancer response to neoadjuvant chemotherapy. We found the combination of these two techniques demonstrated great potential in evaluating the response of breast cancer to neoadjuvant therapy.

Lauren K Fang¹, Ana E Rodríguez-Soto¹, Kathryn E Keenan², and Rebecca A Rakow-Penner^1
1Radiology, University of California San Diego, La Jolla, CA, United States, ²National Institute of Standards and Technology, Boulder, CO, United States

Diffusion-weighted imaging (DWI) has shown potential to diagnose and monitor breast cancer by differentiating benign and malignant lesions. Malignancy-specific cutoff values for DWI estimates have been difficult to establish due to variability across scanners and imaging sites. Thus, understanding the accuracy and precision of these quantitative methods may inform clinical implementation of DW-MRI. Here, we investigated the accuracy and precision of ADC values within a breast phantom across four 3T MRI scanners by the same vendor.

Anabel M Scaranelo¹, Hadassa Degani², Dov Grobgeld³, Vivianne Freitas¹, Shelley Westergard⁴, Christine Elser⁵, and Edna Furman-Haran^3
1Medical Imaging, University of Toronto, Toronto, ON, Canada, ²DDE MRI Solutions Ltd., Tel Aviv, Israel, ³Weizmann Institute of Science, Rehovot, Israel, ⁴Princess Margaret Cancer Center, Toronto, ON, Canada, ⁵University of Toronto, Toronto, ON, Canada

The performance of breast DTI in the diagnosis of breast cancer in high-risk lactating patients as compared to DCE imaging was prospectively assessed. The study included 38 patients of which six were diagnosed with breast cancer. All six cancers were detected by DTI and showed significantly lower diffusion coefficients and lower anisotropy in comparison to contralateral normal breast tissue. Sensitivity, specificity and accuracy were respectively, 100%, 93.2%, 93.8% for DTI and 66.7%, 83.8% and 82.5% for DCE. The results showed the potential of breast DTI as an aid in the diagnostic workup of high-risk women during the lactation period.

Timothy Allen^1,2, Leah C Henze Bancroft², Lloyd Estkowski³, Ty A Cashen³, Frederick Kelcz², Frank R Korosec^1,2, Roberta M Strigel^1,2,4, Orhan Unal^1,2, and James H Holmes^2
1Medical Physics, University of Wisconsin-Madison, Madison, WI, United States, ²Radiology, University of Wisconsin-Madison, Madison, WI, United States, ³Global MR Applications and Workflow, GE Healthcare, Madison, WI, United States, ⁴Carbone Cancer Center, University of Wisconsin-Madison, Madison, WI, United States

A deep learning reconstruction was evaluated for use in T2w breast MRI. Breast radiologists scored deep learning (DL) images significantly higher than non-DL images in four categories: artifacts, perceived signal-to-noise ratio, sharpness, and overall quality. DL was then used to improve the quality of high-resolution T2w breast series acquired in clinically-acceptable scan times. High resolution protocols typically require compromise between scan time and image quality. However, implementation of a deep learning reconstruction allowed for shorter scan times while maintaining diagnostic image quality. A deep learning reconstruction could allow for a clinically-feasible, high-resolution T2w acquisition.   

Muzhen He¹, Huiping Ruan¹, Mingping Ma², Zhongshuai Zhang³, and Robert Grimm^3
1Radiology, Provincial Clinical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, China, ²Provincial Clinical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, China, ³Siemens Healthcare Ltd, Shanghai, China

This study aimed to compare the performance of different diffusion models of，including the mono-exponential, intravoxel incoherent motion (IVIM) model and the diffusion kurtosis imaging (DKI) model，for the correlation with molecular subtype and molecular prognostic factors. The results indicate that ADC, IVIM-D, DKI-D can be used to differentiate luninal B group from others molecular subtype. Quantitative MR diffusion parameters, such as ADC ,DKI-K, DKI-D , IVIM-FP and IVIM-D were significantly associated with HER2 overexpression, Ki-67 proliferation status and histological grade in patients with breast cancer. 

Peiying Zhu¹, Xiaoan Zhang¹, Lin Lu¹, Xin Zhao¹, Qingna Xing¹, Yafei Guo¹, Kaiyu Wang², Jinxia Guo², Xueyuan Wang¹, and Penghua Zhang^1
1Department of Radiology, The Third Affiliated Hospital of Zhengzhou University, Zhengzhou, China, zhengzhou, China, ²GE Healthcare, MR Research China, Beijing, China, zhengzhou, China

Synthetic MRI is able to obtain various image comtrasts and quantitative parameters.These parameters( T1, T2, and PD values) directly reflect the composition of the tissue.The objective of this study was to assess Synthetic magnetic resonance imaging (MRI) ability for differentiation between breast adenosis and breast cancer.Our results show that show that Synthetic MRI is a useful tool that can be utilised to discriminate between breast adenosis and breast cancer.

Mami Iima^1,2, Maya Honda¹, Rie Ota¹, Masako Kataoka¹, Masakazu Toi³, and Yuji Nakamoto^1
1Diagnostic Imaging and Nuclear Medicine, Kyoto University Graduate School of Medicine, Kyoto, Japan, ²Clinical Innovative Medicine, Institute for Advancement of Clinical and Translational Science, Kyoto University Hospital, Kyoto, Japan, ³Breast Surgery, Kyoto University Graduate School of Medicine, Kyoto, Japan

We have 1) assessed the clinical utility of DWI in the diagnosis of breast lesions using DW-based BI-RADS lexicons and lesion conspicuity and 2) investigated the diagnostic performance of breast lesions based on DW-based BI-RADS category by 3 radiologists using breast DW images with b=800 or b=1500 s/mm². Agreement in DW-based BI-RADS categories tended to be higher in b1500 compared to b800 DW images. Diagnostic performance based on b800 and b1500 DW images was not uniform among 3 readers, which highlight the necessity and importance of the standardization on reading protocols and experience from DWI readers.