Review of artificial intelligence technologies for improving the interpretation of diagnostic images in modern neuroradiology

Abstract

Background. The application of artificial intelligence (AI) technologies in neuroradiology is one of the most promising areas of modern medicine. The significant growth in the volume of neuroimaging studies requires automated solutions to improve the accuracy and speed of interpretation. The integration of machine learning (ML) and deep learning (DL) algorithms into clinical practice contributes to the standardization of image analysis and reduces variability in interpretation, which improves the quality of diagnosis and reduces the risk of diagnostic errors. Purpose – to review current research (2022–2025) on the use of AI technologies in neuroradiology, with a focus on ML, DL, hybrid models (ResNet, U-Net, Swin Transformer), and explainable AI (XAI) algorithms. The goal is to evaluate the effectiveness of these technologies in the detection, segmentation, and classification of brain tumors on MRI images. Materials and Methods. A search was conducted in PubMed, Scopus, Web of Science, and EMBASE databases using the keywords: «machine learning», «deep learning», «artificial intelligence», «brain neoplasms», «MRI», «CNN», «U-Net», and «Explainable AI». According to the analysis, 58 articles were included that contained quantitative metrics (accuracy, AUC, Dice), descriptions of architectures, and clinical scenarios. The results are systematized by thematic areas: CNN/U-Net models, hybrid architectures, XAI methods, and clinical integration. Results. Most studies confirm the high effectiveness of AI models in neuroradiology. CNN models provide up to 98.6% accuracy in tumor classification. U-Net and its modifications achieve Dice = 0.97 in the segmentation of neoplasm boundaries. Hybrid architectures (Swin Transformer + ResNet50v2) show up to 99.9% accuracy; ensemble models (DenseNet + InceptionResNetV2) exceed 98–99%. The development of explainable AI increases doctorsʼ confidence in automated decisions. The main barriers are the variability of MRI protocols, the unrepresentativeness of datasets, legal and ethical restrictions. Promising areas include multi-modal analysis, radiogenomics, cloud solutions, and lightweight models for telemedicine. Conclusions. AI technologies are moving from the experimental phase to practical application in neuroradiology. CNN, U-Net, transformer, and hybrid architecture models demonstrate diagnostic accuracy of over 95–99%, which is equal to or exceeds the level of radiologists. Multicenter validation studies, protocol standardization, and regulatory clarification of liability and ethical issues are necessary for widespread clinical implementation.

Ähnliche Arbeiten