Deep Learning-Based Segmentation and Volume Calculation of Pediatric Lymphoma on Contrast-Enhanced Computed Tomographies.

Lymphomas are the ninth most common malignant neoplasms as of 2020 and the most common blood malignancies in the developed world. There are multiple approaches to lymphoma staging and monitoring, but all of the currently available ones, generally based either on 2-dimensional measurements performed on CT scans or metabolic assessment on FDG PET/CT, have some disadvantages, including high inter- and intraobserver variability and lack of clear cut-off points. The aim of this paper was to present a novel approach to fully automated segmentation of thoracic lymphoma in pediatric patients. Manual segmentations of 30 CT scans from 30 different were prepared by the authors. nnU-Net, an open-source deep learning-based segmentation method, was used for the automatic segmentation. The highest Dice score achieved by the model was 0.81 (SD = 0.17) on the test set, which proves the potential feasibility of the method, albeit it must be underlined that studies on larger datasets and featuring external validation are required. The trained model, along with training and test data, is shared publicly to facilitate further research on the topic.

Deep learning for cerebral angiography segmentation from non-contrast computed tomography.

Cerebral computed tomography angiography is a widely available imaging technique that helps in the diagnosis of vascular pathologies. Contrast administration is needed to accurately assess the arteries. On non-contrast computed tomography, arteries are hardly distinguishable from the brain tissue, therefore, radiologists do not consider this imaging modality appropriate for the evaluation of vascular pathologies. There are known contraindications to administering iodinated contrast media, and in these cases, the patient has to undergo another examination to visualize cerebral arteries, such as magnetic resonance angiography. Deep learning for image segmentation has proven to perform well on medical data for a variety of tasks. The aim of this research was to apply deep learning methods to segment cerebral arteries on non-contrast computed tomography scans and consequently, generate angiographies without the need for contrast administration. The dataset for this research included 131 patients who underwent brain non-contrast computed tomography directly followed by computed tomography with contrast administration. Then, the segmentations of arteries were generated and aligned with non-contrast computed tomography scans. A deep learning model based on the U-net architecture was trained to perform the segmentation of blood vessels on non-contrast computed tomography. An evaluation was performed on separate test data, as well as using cross-validation, reaching Dice coefficients of 0.638 and 0.673, respectively. This study proves that deep learning methods can be leveraged to quickly solve problems that are difficult and time-consuming for a human observer, therefore providing physicians with additional information on the patient. To encourage the further development of similar tools, all code used for this research is publicly available.

Automated Ventricular System Segmentation in Paediatric Patients Treated for Hydrocephalus Using Deep Learning Methods.

Hydrocephalus is a common neurological condition that can have traumatic ramifications and can be lethal without treatment. Nowadays, during therapy radiologists have to spend a vast amount of time assessing the volume of cerebrospinal fluid (CSF) by manual segmentation on Computed Tomography (CT) images. Further, some of the segmentations are prone to radiologist bias and high intraobserver variability. To improve this, researchers are exploring methods to automate the process, which would enable faster and more unbiased results. In this study, we propose the application of U-Net convolutional neural network in order to automatically segment CT brain scans for location of CSF. U-Net is a neural network that has proven to be successful for various interdisciplinary segmentation tasks. We optimised training using state of the art methods, including "1cycle" learning rate policy, transfer learning, generalized dice loss function, mixed float precision, self-attention, and data augmentation. Even though the study was performed using a limited amount of data (80 CT images), our experiment has shown near human-level performance. We managed to achieve a 0.917 mean dice score with 0.0352 standard deviation on cross validation across the training data and a 0.9506 mean dice score on a separate test set. To our knowledge, these results are better than any known method for CSF segmentation in hydrocephalic patients, and thus, it is promising for potential practical applications.