The 3D image segmentation is the process of partitioning a digital 3D volumes into multiple segments. This paper presents a fully automatic method for high resolution 3D volumetric segmentation of medical image data using modern supervised deep learning approach. We introduce 3D Dense-U-Net neural network architecture implementing densely connected layers. It has been optimized for graphic process unit accelerated high resolution image processing on currently available hardware (Nvidia GTX 1080ti). The method has been evaluated on MRI brain 3D volumetric dataset and CT thoracic scan dataset for spine segmentation. In contrast with many previous methods, our approach is capable of precise segmentation of the input image data in the original resolution, without any pre-processing of the input image. It can process image data in 3D and has achieved accuracy of 99.72% on MRI brain dataset, which outperformed results achieved by human expert. On lumbar and thoracic vertebrae CT dataset it has achieved the accuracy of 99.80%. The architecture proposed in this paper can also be easily applied to any task already using U-Net network as a segmentation algorithm to enhance its results. Complete source code was released online under open-source license.

This paper proposes an unbalanced end-to-end trained 3D Dense-U-Net network for brain MRI images superresolution. We evaluated capabilites of the proposed architecture on upsampling the MRI brain scans in the factor of 2, 4 and 8 and compared the results with resampled images using lanczos, spline and bilinear interpolation achieving best results. While the network does not exceed superresolution capabilites of state-of-the-art GAN networks, it does not require large dataset, is easy to train and capable of processing 3D images in resolution suitable for medical image processing.

This paper presents a fully automatic method for 3D segmentation of brain tissue on MRI scans using modern deep learning approach and proposes 3D Dense-U-Net neural network architecture using densely connected layers. In contrast with many previous methods, our approach is capable of precise segmentation without any preprocessing of the input image and achieved accuracy 99.70 percent on testing data which outperformed human expert results. The architecture proposed in this paper can also be easily applied to any project already using U-net network as a segmentation algorithm to enhance its results. Implementation was done in Keras on Tensorflow backend and complete source-code was released online.

Diabetic Retinopathy is one of those eye diseases which may cause permanent loss of vision if not treated at an early stage. The current paper proposes an algorithmic rule for detection of red lesions and grading the severity of a fundus image depending on its location in the image. Some significant and deciding objects like optic disc and macula are segmented using adaptive intensity-based threshold, geometrical features, k-means clustering and morphological operations. Imaging techniques like color normalization, median filtering and morphological operations are used for segmentation of blood vessels and red lesions. Finally, a region-based framework has been used for grading the severity of the disease affecting the patient. The proposed method has achieved an accuracy of 89%. The proposed method has given encouraging results and can be used in development of some devices in this field.

Pulmonary fibrosis is one of the most severe long-term consequences of COVID-19. Corticosteroid treatment increases the chances of recovery; unfortunately, it can also have side effects. Therefore, we aimed to develop prediction models for a personalized selection of patients benefiting from corticotherapy. The experiment utilized various algorithms, including Logistic Regression, k-NN, Decision Tree, XGBoost, Random Forest, SVM, MLP, AdaBoost, and LGBM. In addition easily human-interpretable model is presented. All algorithms were trained on a dataset consisting of a total of 281 patients. Every patient conducted an examination at the start and three months after the post-COVID treatment. The examination comprised a physical examination, blood tests, functional lung tests, and an assessment of health state based on X-ray and HRCT. The Decision tree algorithm achieved balanced accuracy (BA) of 73.52%, ROC-AUC of 74.69%, and 71.70% F1 score. Other algorithms achieving high accuracy included Random Forest (BA 70.00%, ROC-AUC 70.62%, 67.92% F1 score) and AdaBoost (BA 70.37%, ROC-AUC 63.58%, 70.18% F1 score). The experiments prove that information obtained during the initiation of the post-COVID-19 treatment can be used to predict whether the patient will benefit from corticotherapy. The presented predictive models can be used by clinicians to make personalized treatment decisions.

This paper compares nine different upsampling methods used in convolutional neural networks in terms of accuracy and processing speed. The process of image segmentation using autoencoder neural networks consists of the image downsampling in the encoder and correspondingly of image upsampling in the decoder part of the network to achieve original image resolution. This paper focuses on the upsampling process in the decoder part of the standard U-Net neural network. Three different interpolations are compared with and without subsequent lxl convolution layers and three transpose convolution layers for image upsampling using different size convolutional cores. The experiment has shown that the best practical results were achieved using simple nearest neighbor interpolation upsampling taking into consideration the computational time needed. The network using nearest neighbor interpolation upsampling achieved pixel accuracy of 99.47% and has shown fast training time and convergence in comparison with other networks using different upsampling methods. The data used in this work consist of a lumbar CT spine segmentation dataset.