Sarah Wacher¹, Jinglei Lv^1,2, and Mariano Cabezas^1
1Brain and Mind Centre, The University of Sydney, Sydney, Australia, ²School of Biomedical Engineering, The University of Sydney, Sydney, Australia

Deep learning approaches for Alzheimer's disease (AD) classification have primarily focused on modelling the structural changes associated with the condition, neglecting the changes in functional brain dynamics. These functional changes may be detectable earlier than structural atrophy providing an avenue for earlier diagnosis and treatment. We therefore proposed a convolutional neural network combined with a long short-term memory unit to decode fMRI signals. The model was able to classify AD from healthy control with a balanced accuracy of 0.69. Whilst there is room to improve network performance, the study already provides promising insights into the possibilities of resting-state fMRI classification.

Lei Wang¹ and Fuqing Zhou^1
1Department of Radiology, The First Affiliated Hospital of Nanchang University, Nanchang, China

Chemotherapy-related cognitive impairment (CRCI), especially subjective cognitive complaints (SCC), had been often reported in breast cancer survivors, which affects their life and work. The identification of biomarkers for early diagnosis and prognosis prediction of SCC remains a crucial challenge of important clinical implications. The resting-state functional magnetic resonance imaging (rs-fMRI) has been widely used to detect abnormalities of brain activity in CRCI. The machine learning method combined with rs-fMRI features could effectively identify breast cancer survivors with chemotherapy-related SCC from healthy controls (HC).

Jin Wang¹, Jiayang Deng², Rui Wang¹, and Jing Zhang^1
1Department of Magnetic Resonance, LanZhou University Second Hospital, Lanzhou, China, ²Vipshop (China) Co. LTD, Guangzhou, China

Isocitrate dehydrogenase (IDH) mutation status has emerged as an important prognostic marker in gliomas.Currently, reliable IDH mutation determination requires invasive surgical procedures. Various studies have demonstrated the efficacy of deep learning in classify IDH status using Magnetic resonance images（MRI）data.In this study we propose a multi-channel architecture of 3D convolutional neural networks (CNNs) based on deep learning to predict the IDH status.We utilize traditional structures imaging and various diffusivity metric maps derived from diffusion tensor imaging (DTI) as input to the network.The final model achieved the AUC value of 0.93.

Bomin Kim¹, Geun Young Lee², and Sung-Hong Park^1
1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Korea, Republic of, ²Deparment of Radiology, Chung-Ang University Hospital, Seoul, Korea, Republic of

Diagnosing stages of osteonecrosis of the femoral head (ONFH) based on MR images can reduce additional cost and radiation exposure caused by CT scan. We propose a deep learning network that enables 4-way classification of ONFH stages, utilizing the information from MR images with different contrasts and planes. Given the limited number of available data, we enhanced the network performance by using self-supervised learning based on MR-to-CT translation task, which increased AUC significantly. We also investigated the diagnostic results from different MR protocols, and obtained more precise and robust results by combining them.

Zihao Tang^1,2, Mariano Cabezas², Kain Kyle^2,3, Arkiev D'Souza², Stephen Tisch⁴, Ben Jonker⁴, Yael Barnett⁴, Joel Maamary⁴, Jerome Maller⁵, Michael Barnett^2,3, Weidong Cai¹, and Chenyu Wang^2,3
1School of Computer Science, University of Sydney, Sydney, Australia, ²Brain and Mind Centre, University of Sydney, Sydney, Australia, ³Sydney Neuroimaging Analysis Centre, Sydney, Australia, ⁴St Vincent's Hospital Sydney, Sydney, Australia, ⁵GE Healthcare Australia, Melbourne, Australia

Disabling tremor is the most common symptom of tremor-dominated Parkinson’s disease (PD) patients. Recently, MR guided focused ultrasound (MRgFUS) has been applied in the clinical environment to treat tremor. However, patients can have different treatment responses to the ablation. Tremor suppression can be observed on some patients after ablating the ventral intermediate nucleus (Vim), while others require further ablations. To provide a tailored treatment, a deep learning technique is introduced in this work to predict whether a subject undergoing MRgFUS will respond to a single ablation in the VIM or require additional lesioning in other regions.

Netanell Avisdris^1,2, Daphna Link Sourani¹, Liat Ben Sira^3,4,5, Leo Joskowicz², Gustavo Malinger^4,6, Simcha Yagel⁷, Elka Miller⁸, and Dafna Ben Bashat^1,4,5
1Sagol Brain Institute, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, ²School of Computer Science and Engineering, The Hebrew University of Jerusalem, Jerusalem, Israel, ³Division of Pediatric Radiology, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, ⁴Sackler Faculty of Medicine, Tel Aviv University, Tel Aviv, Israel, ⁵Sagol School of Neuroscience, Tel Aviv University, Tel Aviv, Israel, ⁶Division of Ultrasound in Obstetrics & Gynecology, Lis Maternity Hospital, Tel Aviv, Israel, ⁷Obstetrics and Gynecology Division, Hadassah Hebrew University Medical Center, Jerusalem, Israel, ⁸Medical Imaging, CHEO, University of Ottawa, Ottawa, ON, Canada

Brain MRI of 296 fetuses (22-40 weeks’ gestational-age, normal n=244, hypo/hyper-telorism n=52 were included. Binocular, interocular, and ocular diameters, and ocular volume were measured using automatic methods. Two new parameters, binocular-ratio and interocular-ratio, were defined. In normal fetuses, all four measurements increased with gestational-age. However, constant values were detected across all gestational-ages of binocular-ratio=1.56±0.05 and interocular-ratio=0.60±0.05. A random-forest classifier achieved the best results (out of eight classifiers) with AUC-ROC=0.90±0.03 for classification between normal and fetuses with hypo/hyper-telorism. mainly based on the two new ratios. Machine-learning multi-parametric classification and the new ratios are suggested to be used in routine clinical practice.

Lucie Piram¹, Acquitter Clément², Julia Gilhodes³, Umberto Sabatini⁴, Elizabeth Cohen-Jonathan Moyal^1,5, Benjamin Lemasson², and Soleakhena Ken^5,6,7
1Department of Radiotherapy, Institut Universitaire du Cancer de Toulouse - Oncopole, Toulouse, France, ²Grenoble Institut des Neurosciences, Grenoble, France, ³Department of Clinical Trials, Institut Universitaire du Cancer de Toulouse - Oncopole, Toulouse, France, ⁴Dipartimento di Scienze Mediche e Chirurgiche, Università Magna Graecia, Catanzaro, Italy, ⁵U1037, RADOPT Team, Cancer Research Center of Toulouse, Toulouse, France, ⁶Department of Engineering and Medical Physics, Institut Universitaire du Cancer de Toulouse - Oncopole, Toulouse, France, ⁷MINDS Team UMR 5505, Institut de Recherche en Informatique de Toulouse, Toulouse, France

Radiomic features computed from multiparametric MRI were found to be relevant as early in-vivo biomarkers for pseudoprogression evaluation in recurrent glioblastoma patients. At baseline, predictive biomarker for pseudoprogression outcome was related to kurtosis parameter of FLAIR histogram plotted from abnormal hyper-intense signal area. When considering variation between baseline and first event (either pseudoprogression or true progression), four early biomarkers were found for entropy of T1-weighted, T1-weighted-post-contrast morphological MRI and Apparent Diffusion Coefficient maps derived from diffusion-weighted MRI.

Such early in-vivo biomarkers easily computed from automatic segmentation and first order radiomics analysis could be useful for the assessment of treatment response.

Seyyed Ali Hosseini¹, Isaac Shiri², Ghasem Hajianfar³, Stephen Bagley⁴, MacLean Nasrallah⁵, Donald M O’Rourke ⁶, Suyash Mohan⁷, and SANJEEV CHAWLA^7
1Medical physics and biomedical engineering, Tehran university of medical science, Karaj, Iran (Islamic Republic of), ²Medical Imaging, University of Geneva, Geneva, Switzerland, ³Cardiovascular Medical and Research Center, Iran University of Medical Science, Tehran, Iran (Islamic Republic of), ⁴Hematology-Oncology, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA, United States, ⁵Pathology and Lab Medicine, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA, United States, ⁶Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA, United States, ⁷Radiology, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA, United States

To distinguish IDH-mutant from IDH wild-type grade-4 astrocytomas, conventional MR imaging (post-contrast T1-weighted and T2-Flair images) was acquired from 57 patients [IDH-mutant (n=23) and IDH-wild-type (n=34) grade-4 astrocytomas]. Post-contrast T1-weighted and T2-FLAIR images were resliced, resampled and co-registered. Neoplasms were segmented into whole tumor, enhancing region, central necrotic region, edema region, and core tumor (contrast enhancing + necrotic region). A total of 105 first-, second-, higher-order and shape based radiomics features were extracted from each ROI. Readily interpretable and quantitative features from different sub-regions of neoplasms were observed with high diagnostic performances in distinguishing IDH-mutant from IDH wild-type grade-4 astrocytomas.

Vishwanatha Mitnala Rao¹, Junhao Zhang¹, and Jia Guo^2
1Department of Biomedical Engineering, Columbia University, New York, NY, United States, ²Department of Psychiatry, Columbia University, New York, NY, United States

Schizophrenia diagnosis is clinically difficult due to the lack of biomarkers associated with the disease. While machine learning algorithms and convolutional neural networks (CNNs) have found success using neuroimaging inputs to diagnose the disease, they have historically not performed or generalized well enough for clinical use. We propose 3D-MIC-Transformer, the first transformer-based deep learning architecture applied to neurological disease classification that demonstrates state-of-the-art schizophrenia classification performance and generalization using structural MRI inputs. 3D-MIC-Transformer outperforms prior CNN implementations (AUROC: 0.985, accuracy: 0.933), and we believe 3D-MIC-Transformer can serve as a backbone for other disease classification tasks in the future.

Buse Buz-Yalug¹, Ayca Ersen Danyeli^2,3, Kubra Tan⁴, Ozge Can⁵, Necmettin Pamir^3,6, Alp Dincer^3,7, Koray Ozduman^3,8, and Esin Ozturk-Isik^1
1Institute of Biomedical Engineering, Bogazici University, Istanbul, Turkey, ²Department of Medical Pathology, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ³Center for Neuroradiological Applications and Reseach, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁴Health Institutes of Turkey, Istanbul, Turkey, ⁵Department of Medical Engineering, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁶Department of Neurosurgery, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁷Department of Radiology, Acıbadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁸Department of Neurosurgery, Acıbadem Mehmet Ali Aydinlar University, Istanbul, Turkey

Meningiomas are the most frequent primary intracranial and spinal tumors. In this work, we studied the relative cerebral blood volume (rCBV) changes in the meningiomas with NF2 loss and S100 positivity. Meningiomas with NF2-L had lower rCBV than NF2-NL, and S100 positive group had also lower rCBV than S100 negative group. The highest classification accuracies obtained using machine learning applications were 75.6% for NF2 molecular subsets and 75.2% for S100 molecular subsets.

Anish Raj¹, Fabian Tollens², Anika Strittmatter¹, Laura Hansen¹, Dominik Noerenberg², and Frank G Zöllner^1
1Computer Assisted Clinical Medicine, Mannheim Institute for Intelligent Systems in Medicine, Medical Faculty Mannheim, Heidelberg University, Mannheim, Germany, ²Department of Clinical Radiology and Nuclear Medicine, Medical University Center Mannheim, Medical Faculty Mannheim, Heidelberg University, Mannheim, Germany

The total kidney volume (TKV) increases with ADPKD progression and hence, can be used to quantify disease progression. The TKV calculation requires accurate delineation of kidney volumes which is usually performed manually by an expert physician. However, this is time consuming and automated segmentation is warranted, e.g., using deep learning. The implementation of the latter is usually hindered due to a lack of large, annotated datasets.In this work, we address this problem by implementing the cosine loss function and a technique called Sharpness Aware Minimization (SAM) into the U-Net to improve TKV estimation in small sized datasets.