Analysis of deep learning approaches for automated prostate segmentation: literature review

Le résumé fourni par la source

Background. Delineation of the prostate boundaries represents the initial step in understanding the state of the whole organ and is mainly manually performed, which takes a long time and directly depends on the experience of the radiologists. Automated prostate selection can be carried out by various approaches, including using artificial intelligence and its subdisciplines – machine and deep learning.Aim. To reveal the most accurate deep learning-based methods for prostate segmentation on multiparametric magnetic resonance images.Materials and methods. The search was conducted in July 2022 in the PubMed database with a special clinical query (((AI) OR (machine learning)) OR (deep learning)) AND (prostate) AND (MRI). The inclusion criteria were availability of the full article, publication date no more than five years prior to the time of the search, availability of a quantitative assessment of the reconstruction accuracy by the Dice similarity coefficient (DSC) calculation.Results. The search returned 521 articles, but only 24 papers including descriptions of 33 different deep learning networks for prostate segmentation were selected for the final review. The median number of cases included for artificial intelligence training was 100 with a range from 25 to 365. The optimal DSC value threshold (0.9), in which automated segmentation is only slightly inferior to manual delineation, was achieved in 21 studies.Conclusion. Despite significant achievements in the development of deep learning-based prostate segmentation algorithms, there are still problems and limitations that should be resolved before artificial intelligence can be implemented in clinical practice.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.