Evgeniy Pavlovskiy

Исследование направлено на анализ современных алгоритмов для диагностики поражений при рассеянном склерозе на МРТ-изображениях. Рассеянный склероз, тяжелое заболевание центральной нервной системы, занимает первое место среди причин инвалидности у пациентов молодого трудоспособного возраста. В связи с развитием технологий компьютерного зрения и машинного обучения растет актуальность применения данных технологий для медицинской диагностики. Такие подходы необходимы для эффективной разработки и внедрения диагностических систем с использованием искусственного интеллекта. Материал и методы. В статье представлены особенности диагностики рассеянного склероза на МРТ-изображениях, существующие наборы данных: ISBI-2015, MSSEG-2016, MSSEG-2021; существующие алгоритмы и модели сегментации поражений: U-Net, nnU-Net, TransUnet, TransBTS, UNETR, Swin UNETR. Результаты и их обсуждение. Проведено обучение и сравнение архитектур и моделей nnU-Net, UNETR, Swin UNETR на ISBI-2015 c различными параметрами и функциями потерь, использованы четыре последовательности МРТ: Т2-взвешенное изображение, T2-FLAIR, PD, MPRAGE. Сегментация поражений одобрена тремя аттестованными опытными нейрорадиологами. Заключение. Описанные в статье подходы, включая процессы обработки данных, обучения моделей, анализ результатов, были сосредоточены на выборе и разработке высококачественных алгоритмов компьютерного зрения для определения поражений …

—Одной из ключевых задач Экспертного совета Академпарка является задача экспертизы проектов компаний, претендующих на статус резидентов Технопарка. Необходимо оценить научно-технический и/или технологический уровень проекта, его инновационность, перспективы коммерциализации, квалификацию команды проекта и другие составляющие. В составе Экспертного совета Технопарка 30 специалистов из СО РАН, НГУ, компаний-резидентов Технопарка и государственных структур—пояснил Сергей Кобцев.Новосибирский государственный университет является обладателем уникального ресурса—высококвалифицированных исследователей и разработчиков, имеющих научные компетенции самого высокого уровня. Профессиональные компетенции сотрудников НГУ широко востребованы для качественной и глубинной оценки научных результатов …

The presented paper is devoted to the numerical study of the applicability of 3D inversion for fracture model reconstruction based on machine learning. In practice, geophysicists use seismic inversion for predicting reservoir properties. One-dimensional convolutional model lies in the basis of standard versions of inversion, but geology is more complex. That is why we provide implementation and investigation of the approach for 3D fracture model reconstruction based machine learning, which uses U-net neural network and 3D convolutional model. We provide numerical results for a realistic 3D synthetic fractured model from the North of Russia.

The presented paper is devoted to the numerical study of the applicability of 1D seismic inversion and 2D machine learning based inversion for fracture model reconstruction. Seismic inversion is used to predict reservoir properties. Standard version is based on a one-dimensional convolutional model, but real geological media are more complex, and therefore it is necessary to determine conditions where seismic inversion gives acceptable results. For this purpose, the work carries out a comparative analysis of one-dimensional and two-dimensional convolutional modeling. Also, machine learning methods have been adopted for 2D fracture model reconstruction. We use UNet architecture and 2D convolutional model to create a training dataset. We perform numerical experiments for a realistic synthetic model from Eastern Siberia and Sigsbee model.

BACKGROUND: The main reason for the development and implementation of artificial intelligence (AI) technologies in neuro-oncology is the high prevalence of brain tumors reaching up to 200 cases per 100,000 population. The incidence of a primary focus in the brain is 5%–10%; however, 60%–70% of those who die from malignant neoplasms have metastases in the brain. Magnetic resonance imaging (MRI) is the most common method for primary non-invasive diagnosis of brain tumors and monitoring disease progression. One of the challenges is the classification of tumor types and determination of clinical parameters (size and volume) for the conduct, diagnosis, and treatment procedures, including surgery.AIM: To develope a software module for the differential diagnosis of brain neoplasms on MRI images.METHODS: The software module is based on the developed Siberian Brain Tumor Dataset (SBT), which contains information on over 1000 neurosurgical patients with fully verified (histologically and immunohistochemically) postoperative diagnoses. The data for research and development was presented by the Federal Neurosurgical Center (Novosibirsk). The module uses two-and three-dimensional computer vision models with pre-processed MRI sequence data included in the following packages: pre-contrast T1-weighted image (WI), post-contrast T1-WI, T2-WI, and T2-WI with fluid-attenuated inversion-recovery technique. The models allow to detect and recognize with high accuracy 4 types of neoplasms, such as meningioma, neurinoma, glioblastoma, and astrocytoma, and segment and distinguish components and sizes: ET (tumor …

The study of brain tumor structure and its type-dependent variations is one of the most important research areas in which medical imaging techniques are used. The structural and statistical analysis of these lesions raises various related problems and projects, such as the detection of the neuro oncology diseases, the shape and the segmentation of specific sub-regions (i.e. necrotic part, (non-)enhanced part, edema), the classification of the tumor occurrence and the subsequent treatment up-prognosis. Almost all of these problems are usually solved numerically, particularly with the tendency to use methods related to artificial intelligence (AI), often including deep learning (DL) networks. One of the most complicated, least researched and challenging tasks in this field is the classification of tumor types. This difficulty can be explained by several reasons, the most important of which is the severe limitation of existing …

Histopathological examination of biopsy tissues is still utilized to diagnose and classify brain cancers today. The current approach is inconvenient, time-consuming, and prone to human mistake. These disadvantages emphasize the significance of establishing a fully automated deep learning-based system for classifying brain tumors. In this paper, we suggest an approach to improve the classification for four types of brain tumors by providing the classifier with segmentation as semantic features. 1,452 multi model magnetic resonance images from the Siberian Brain Tumor Dataset (SBT) are used for training, validation, and testing. The training and validation are implemented with our experimental simple convolutional neural network and a pre-trained VGG16. Best performed models are selected and tested on both SBT and the Brain Tumor Segmentation Challenge 2020 dataset (BraTS). The models with …

Recently, the advancement in quantum technologies has had massive impact on the development of quantum algorithms on near-term quantum devices. Variational circuits, a combination of both quantum and classical algorithms, have been very useful in this advancement on near-term quantum devices. Despite these advances, most quantum applications in machine learning (deep learning) especially in transfer learning have been proof-of-concept in the qubit system and very little in the continuous-variable space but no or little application to audio data. This study applies variational circuits to practical real-life speaker classification data for the first time in the continuous-variable system. In separate experiments, the quantum model was combined with a simple convolutional neural network and ResNet18 model, respectively, and the results were compared to classical ResNet18 model applied on the …

Brain tumor segmentation is an important and time-consuming part of the usual clinical diagnosis process. Multi-class segmentation of different tumor types is a challenging task, due to the differences in shape, size, location and scanner parameters. Many 2D and 3D convolution neural network architectures have been proposed to address this problem achieving a significant success. It is well known that 2D approach is generally faster and more popular in the most of such problems. However, the usage of 3D models allows us to simultaneously improve the quality of segmentation. Accounting the context along the sagittal plane leads to the learning of 3-dimensional features that we used for computationally expensive 3D operations what in its turn increases the learning time as well as decreases the speed of operation.In this paper, we compare the 2D and 3D approaches on 2 datasets with MRI images: the one …

Исследование направлено на анализ современных подходов к организации и методологии проектирования базы данных визуализации, построенной на основе компьютерного зрения. Такие подходы необходимы для эффективной разработки диагностических систем с использованием искусственного интеллекта (ИИ). Обязательным условием для этого является качественный набор обучающих данных. Материал и методы. В статье представлена технология создания аннотированной базы данных (SBT Dataset), содержащей около 1000 клинических случаев на основе архивных данных ФГБУ «Федеральный нейрохирургический центр», Новосибирск, Россия, включая сведения о пациентах с астроцитомой, глиобластомой, менингиомой, невриномой и больных с метастазами соматических опухолей. Каждый случай представлен предоперационной МРТ …

Исследование направлено на анализ современных подходов к организации и методологии проектирования базы данных визуализации, построенной на основе компьютерного зрения. Такие подходы необходимы для эффективной разработки диагностических систем с использованием искусственного интеллекта (ИИ). Обязательным условием для этого является качественный набор обучающих данных. Материал и методы. В статье представлена технология создания аннотированной базы данных (SBT Dataset), содержащей около 1000 клинических случаев на основе архивных данных ФГБУ «Федеральный нейрохирургический центр», Новосибирск, Россия, включая сведения о пациентах с астроцитомой, глиобластомой, менингиомой, невриномой и больных с метастазами соматических опухолей. Каждый случай представлен предоперационной МРТ. Результаты и их обсуждение. Построен набор данных (набор данных SBT), содержащий сегментированные 3D МРТ-изображения 5 типов опухолей головного мозга с общим количеством проверенных наблюдений 991. Использованы четыре последовательности МРТ - T1-WI, T1C (с Gd-контрастом), T2-WI и T2-FLAIR с гистологическим и гистохимическим послеоперационным подтверждением. Сегментация опухолей с проверкой границ элементов ядра опухоли и перифокального отека одобрена двумя аттестованными опытными нейрорадиологами. Вывод. База данных, построенная в ходе исследования, по своему объему и уровню качества (верификации) сравнима с современными наиболее …

Программа предназначена для использования при дооперационной МРТ-диагностике в нейрохирургии. Позволяет обнаруживать и распознавать 4 типа опухолей головного мозга: менингиома, невринома, глиобластома, астроцитома, а также сегментировать и выделять компоненты и размеры опухоли: некроз, поглощающая Gd-контраст опухолевая ткань, непоглощающая Gd-контраст опухолевая ткань, перитуморальный отек. Программный модуль основан на 3-мерной модели компьютерного зрения, с предварительной обработкой данных МРТ-последовательности, включенные в пакеты: предконтрастное, постконтрастное Т1-взвешенное изображение, T2-взвешенные изображения с технологией инверсии-восстановления с ослаблением сигнала от жидкости. ОС: Ubuntu 18.04 LTS.

In structural and statistical analysis of neurooncological lesions, there are problems of detection, classification, contouring of tumors, and segmentation of its subregions (Gdcontrast-accumulating, Gd-contrast-non-accumulating, necrotic part of the tumor, perifocal edema), treatment prognosis. Almost all of these tasks are usually solved diagnosis neuro-oncology diseases with use the Artificial Intelligence (AI) applications. AI models training is based on the" Federal Neurosurgical Center" in Novosibirsk, Russia data on patients, who underwent MRI imaging of the brain on a 1.5 T (Siemens Magnetom Avanto) and 3.0 T (Ingenia, Philips) magnetic resonance tomographs: astrocytomas (70; 39% male; median= 37, IQR (31; 56) years), glioblastomas (140; 53% male; median= 57, IQR (48; 63) years), meningiomas (160; 26% male; median== 58, IQR (50; 64) years), neurinomas (130; 32% male; median= 53, IQR (40; 59) years). MRI sequences: T1-WI, T1-WI-Gd+, T2-WI, T2-FLAIR and DWI; scans in a single isovoxel anatomical template with a resolution of 1 mm3. The data were processed in the open-source 3D Slicer (www. slicer. org) cross-platform system with manual tumor segmentation and post-correction by two expert neuroradiologists. Firstly, we elaborated 2D approach for segmentation and classification tasks. As part of the PyTorch machine learning framework, two convolutional neural networks have been developed based on the se-resnext50-32x4d, se-resnext101-32x4d networks [1]. Segmentation of tomograms is implemented according to a multilayer model of a brain tumor (layer= disease) of logically connected 4 blocks (= variant or …

In this study, we provide segmentations of brain tumors as semantic features to a simple convolutional neural network (CNN) to improve the classification results. The Siberian Brain Tumor Dataset (SBT) of 1452 magnetic resonance (MR) images of Russian people’s brains is used for training, validation, and testing. We preprocess MR images by removing the swelling region surrounding a brain tumor (edema) from the segmentation and eliminating slices that contain less than 163 pixels of tumor. The binary classifier is a simple network of four convolutional layers for feature extractions and two linear layers for classification. The network is trained and validated with 5-fold cross-validation. We train another CNN with similar configuration on images without provided segmentation and compare the testing results. The two networks are evaluated with four metrics: accuracy, sensitivity, specificity, and F1 scores. The …

In the medical industry, misdiagnosis of disease is acknowledged as the most common and harmful medical errors as it can cost a human life. Radiologists require a lot of time to manually annotate and segment the images. Over the several years, deep learning has been playing a vital role in the field of computer vision. One of its key uses in the medical industry is to minimize misdiagnosis and the amount of time taken to annotate and segment the images. In this paper, a new deep learning approach for brain tumor classification on MRI Images is introduced. A deep neural network is pretrained as a discriminator in a generative adversarial network (GAN) on MR Images by using multi-scale gradient GAN (MSGGAN) with auxiliary classification to extract the features and to classify the images. In the discriminator, one of the fully connected blocks acts as an auxiliary classifier and the other fully connected block acts as …

Neuro-oncological MRI imaging is a complex, expensive procedure that is responsible for all further treatment tactics. The following issues must be unambiguously resolved: (1) to detect a volumetric process in the brain (e.g., tumor); (2) to outline the exact boundaries of the tumor (to delimit the edematous zone and healthy brain tissue); (3) to determine the level of tumor malignancy as accurately as possible. Artificial intelligence technologies make it possible to speed up the process of MRI diagnostics via 3D visualization and increase its accuracy and specificity. This paper presents pipeline and approaches to the creation of a dataset, which can serve as a basis for solving the problems mentioned above. The description of the dataset which is formed in our research project is presented. The methods and algorithms that were used to solve the problem of multiclass segmentation of the tumor are also described.

Investigation of brain tumor structure and its type-dependent variations are among the list of most important research directions where the medical imaging methods are used. Structural and statistical analysis of these lesions originates various associated problems and projects such as detection of the tumors, shape and specific sub-regions segmentation (i.e. necrotic part, (non-)enhanced part, edema), classification of the tumor presence and treatment follow up prognosis. Almost all of these problems are usually solved numerically, specifically with the tendency to use the Artificial Intelligence (AI) related methods often including Deep Learning (DL) networks. One of the most complicated, weakly explored and challenging tasks in this domain is the classification of the tumor types. This difficulty is explained by several reason where the most principle one is the strong limitation of the existing open-sourced datasets that include clinically confirmed tumor type labels based on the radiological examination protocols. In this work we present current results of the brain tumor classification problem, where we consider and operate with four different lesion types such as meningioma, neurinoma, glioblastoma and astrocytoma. All the conducted research and presented results are obtained on the newly introduced dataset including 255 labeled volume MRI scans describing wide variety of the tumors and its clinically associated ground truth (GT) information. Obtained in this work results demonstrate not only inspiring and strong Accuracy performance of 0.925 on patient level (and accordingly 0.894 slice-wise) but also very high potential and perspective for …

The brain tumor segmentation is one of the crucial tasks nowadays among other directions and domains where daily clinical workflow requires to put a lot of efforts while studying computer tomography (CT) or structural magnetic resonance imaging (MRI) scans of patients with various pathologies. MRI is the most common method of primary detection, non-invasive diagnostics and a source of recommendations for further treatment of brain diseases. The brain is a complex structure, different areas of which have different functional significance. In this paper, we extend the previous research work on the robust pre-processing methods which allow to consider all available information from MRI scans by the composition of T1, T1C, T2 and T2-Flair sequences in the unique input. Such approach enriches the input data for the segmentation process and helps to improve the accuracy of the segmentation and associated …

We apply a cascaded training pipeline for the 3D U-Net to segment each brain tumor sub-region separately and chronologically. Firstly, the volumetric data of four modalities are used to segment the whole tumor in the first round of training. Then, our model combines the whole tumor segmentation with the mpMRI images to segment the tumor core. Finally, the network uses whole tumor and tumor core segmentations to predict enhancing tumor regions. Unlike the standard 3D U-Net, we use Group Normalization and Randomized Leaky Rectified Linear Unit in the encoding and decoding blocks. We achieved dice scores on the validation set of 88.84, 81.97, and 75.02 for whole tumor, tumor core, and enhancing tumor, respectively.

In this paper, we extend the previous research works on the robust multi-sequences segmentation methods which allows to consider all available information from MRI scans by the composition of T1, T1C, T2 and T2-FLAIR sequences. It is based on the clinical radiology hypothesis and presents an efficient approach to combining and matching 3D methods to search for areas of comprised the GD-enhancing tumor in order to significantly improve the model’s performance of the particular applied numerical problem of brain tumor segmentation.Proposed in this paper method also demonstrates strong improvement on the segmentation problem. This conclusion was done with respect to Dice and Hausdorff metric, Sensitivity and Specificity compare to identical training/test procedure based only on any single sequence and regardless of the chosen neural network architecture. We achieved on the test set of 0.866, 0 …