Lesion segmentation in lung CT scans using unsupervised adversarial learning

• DOI: 10.1007/s11517-022-02651-8
• Publication Date: 2022-09-20T09:04:03Z
• Authors: Moiz Khan Sherwani; Aldo Marzullo; Elena De Momi; Francesco Calimeri
• Keywords: Human physiology

Abstract Lesion segmentation in medical images is difficult yet crucial for proper diagnosis and treatment. Identifying lesions costly time-consuming requires highly specialized knowledge. For this reason, supervised semi-supervised learning techniques have been developed. Nevertheless, the lack of annotated data, which common imaging, an issue; context, interesting approaches can use unsupervised to accurately distinguish between healthy tissues lesions, training network without using annotations. In work, technique proposed automatically segment coronavirus disease 2019 (COVID-19) on 2D axial CT lung slices. The approach uses image translation generate based infected need lesion Attention masks are used improve quality further. Experiments showed capability it outperforms a range detection approaches. average reported results test dataset metrics: Dice Score, Sensitivity, Specificity, Structure Measure, Enhanced-Alignment Mean Absolute Error 0.695, 0.694, 0.961, 0.791, 0.875, 0.082 respectively. achieved promising compared with state-of-the-art could constitute valuable tool future developments. Graphical abstract