Angelo Zuffianò¹, Bob de Vos¹, Jorrit Glastra¹, Pim Moeskops¹, Valerio Fortunati¹, Ivana Išgum², Tim Leiner³, Carla van Gils³, and Wouter Veldhuis^3
1Quantib, Utrecht, Netherlands, ²Biomedical Engineering and Physics, Radiology and Nuclear Medicine, Amsterdam University Medical Center, Amsterdam, Netherlands, ³Radiology, Utrecht University Medical Center, Utrecht, Netherlands

Dynamic contrast enhanced (DCE) MRI is the key series to analyze for the detection of breast cancer in women with extremely dense breasts. Given the increasing number of women receiving dense breast MRI screening we aimed to reduce radiologist workload without reducing the high sensitivity of MRI. We developed a convolutional neural network (CNN) based method able to defer 8.1% of the workload by identifying non-enhancing scans with a sensitivity of 96.3%.

Artur Hahn^1,2, Julia Krüwel-Bode³, Yannis Seemann², Sarah Schuhegger², Johann M. E. Jende¹, Anja Hohmann⁴, Volker J. F. Sturm¹, Ke Zhang⁵, Sabine Heiland¹, Martin Bendszus¹, Michael O. Breckwoldt^1,6, Christian H. Ziener^1,5, and Felix T. Kurz^1,5
1Department of Neuroradiology, Heidelberg University Hospital, Heidelberg, Germany, ²Department of Physics and Astronomy, University of Heidelberg, Heidelberg, Germany, ³Molecular Mechanisms of Tumor Invasion (V077), German Cancer Research Center (DKFZ), Heidelberg, Germany, ⁴Department of Neurology, Heidelberg University Hospital, Heidelberg, Germany, ⁵Department of Radiology (E010), German Cancer Research Center (DKFZ), Heidelberg, Germany, ⁶Clinical Cooperation Unit Neuroimmunology and Brain Tumor Immunology, German Cancer Research Center (DKFZ), Heidelberg, Germany

Based on the microvasculature of entire healthy and tumor-bearing mouse brains, imaged with high-resolution fluorescence microscopy, the transverse relaxation process within virtual MRI voxels was simulated. Extended parametrizations of the non-Lorentzian signal decay were used to train support vector machine and random forest classifiers to differentiate healthy brain and tumor voxel signals. A proof-of-principle is presented with U87 and GL261 glioblastoma at different SNR levels. This automated workflow enables the in-silico development of specialized MRI sequences to maximize classification accuracy with minimal NMR measurements for experimental analogies.

Benjamin W Wormald^1,2, Thomas EJ Ind^2,3, and Nandita M deSouza^1,4
1Imaging, The Institute of Cancer Research, Sutton, Surrey, United Kingdom, ²Gynaecological Oncology, The Royal Marsden NHS FoundationTrust, London, United Kingdom, ³Surgery, St. Georges University Hospital, London, United Kingdom, ⁴MRI Unit, The Royal Marsden NHS Foundation Trust, Sutton, Surrey, United Kingdom

Cervical cancer recurs post-trachelectomy often because of close surgical margins or lymph-node micrometastases. We show that 5 texture features distinguish good- from poor-prognosis tumors (low/high volume, without/with parametrial invasion, without/with lymph node metastases).  For tumors suitable for trachelectomy (<4.19cm³), linear regression of feature value with volume (using 3 features with high discrimination of groups and 1 standard deviation from median from good prognosis group as threshold) indicated that radiomic features tended towards values representing poor prognosis at 1.8±0.2cm³ (T2-W images) and 1.8±0.06cm³ (ADC maps). Above 1.8cm³ textural features of cervical cancer shift towards a phenotype likely to spread and metastasize.

Asim M. Mazin¹, Samuel H. Hawkins¹, Olya Stringfield², Jasreman Dhillon³, Brandon J. Manley⁴, Daniel K. Jeong⁵, and Natarajan Raghunand^1
1Department of Cancer Physiology, Moffitt Cancer Center, Tampa, FL, United States, ²IRAT Shared Service, Moffitt Cancer Center, Tampa, FL, United States, ³Department of Anatomic Pathology, Moffitt Cancer Center, Tampa, FL, United States, ⁴Department of Genitourinary Oncology, Moffitt Cancer Center, Tampa, FL, United States, ⁵Department of Diagnostic & Interventional Radiology, Moffitt Cancer Center, Tampa, FL, United States

We report a machine learning approach using Self-Organizing Maps (SOM) and Learning Vector Quantization (LVQ) to analyze multiparametric MRI for the purpose of differentiating between renal cell carcinoma tumor with (sRCC) and without (nsRCC) sarcomatoid de-differentiation, a transformation that is associated with poorer outcomes. The SOM+LVQ model was trained on mpMRI data from 9 nsRCC and 9 sRCC tumors, validated on a separate cohort of 3 nsRCC and 3 sRCC tumors, and tested on a held-out set of 5 nsRCC and 5 sRCC tumors. An overall classification accuracy of 70% was achieved on the test cohort.

Hongyi Duanmu¹, Pauline Huang¹, Srinidhi Brahmavar¹, Fusheng Wang¹, and Tim Q Duong^1
1Stony Brook University, Stony Brook, NY, United States

Pathological complete response (pCR) is a measurement of the effectiveness of neoadjuvant chemotherapy (NAC). While there are several studies about predicting the pCR, no one system can fully automate this prediction process. We proposed a 3D convolutional neural network (CNN) system, integrating information on breast MRI images and prognostic clinical features, to predict pCR pre-NAC. This system achieved inspiring results in the ISPY1 Clinical Trial dataset, with 77% accuracy. This approach shows the potential in breast cancer diagnose and assessment. Furthermore, the mechanism of integrating images and features information can be used and generalized to other tasks.

James C Korte^1,2, Carlos E Cardenas³, Tomas Kron^1,4, Nicholas Hardcastle^1,5, Jihong Wang³, Houda Bahig⁶, Baher Elgohari⁷, Laurence E Court³, Clifton D Fuller⁷, and Sweet Ping Ng^7,8
1Department of Physical Science, Peter MacCallum Cancer Centre, Melbourne, Australia, ²Department of Biomedical Engineering, The University of Melbourne, Melbourne, Australia, ³Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, ⁴Sir Peter MacCallum Department of Oncology, The University of Melbourne, Melbourne, Australia, ⁵Centre for Medical Radiation Physics, University of Wollongong, Wollongong, Australia, ⁶Radiation Oncology Department, Centre Hospitalier de l'Université de Montréal, Montreal, QC, Canada, ⁷Department of Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, TX, United States, ⁸Department of Radiation Oncology, Peter MacCallum Cancer Centre, Melbourne, Australia

Radiomics is a promising technique for discovering image based biomarkers of therapy response in cancer. Reproducibility of radiomic features is a known issue that is being addressed by standardisation initiatives, but it remains a challenge to interpret previously published radiomic signatures. We investigate the reproducibility of radiomic features calculated with two common software packages and explore the impact of including non-reproducible diffusion features in a head and neck cancer (HNC) radiotherapy response model. Our results demonstrate that equivalent models can be generated from either software, but only when restricting the model to reproducible features identified with a correlation threshold method.

Lars Bielak^1,2, Nicole Wiedenmann^2,3, Arnie Berlin⁴, Leonard Hägele¹, Thomas Lottner¹, Sebastian Gross⁵, Anca-Ligia Grosu^2,3, and Michael Bock^1,2
1Dept. of Radiology, Medical Physics, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany, ²German Cancer Consortium (DKTK), Partner Site Freiburg, Freiburg, Germany, ³Dept. of Radiation Oncology, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany, ⁴The MathWorks, Inc., Novi, MI, United States, ⁵The MathWorks, Inc., Ismaning, Germany

Multiparametric MRI imaging in combination with PET/CT is the basis for precise tumor segmentation in radiation therapy. We trained a segmentation CNN on the multiparametric MRI data of head and neck squamous cell carcinoma patients and investigated the network robustness against noise corruption in the input channels. Overall noise robustness and differences between seven different input contrasts were compared.

Xiao-li Song¹, Jia-Liang Ren², Kaiyu Wang³, and JinLiang Niu^1
1Shanxi Medical University Second Affiliated Hospital, Taiyuan, China, ²GE Healthcare, Beijing, China., Beijing, China, ³GE Healthcare, MR Research China, Beijing, China, Beijing, China

DCE-MRI and its subsequently derived pharmacokinetic parameters have been adopted to explore tumor angiogenesis and vascular permeability changes inside tumors and improve the diagnostic accuracy of ovarian tumors. Radiomics can convert medical images to mineable high-dimensional quantitative imaging features based on automatic feature extraction algorithms. In this study, we present a radiomics model based on a DCE-MRI PK protocol and establish an effective and noninvasive 3-class classification prediction model for the discrimination among benign, borderline and malignant ovarian tumors.