This paper presents a new automated unsupervised segmentation system to accurately delineate the pulmonary region in 3D computed tomography (CT) scans. It operates on a multi-dimensional joint probability model, which leverages a deep learning-based transformer model to optimize the log-likelihood of that probabilistic model. This probability model integrates appearance probability models, that represent various radiodensity distributions in both lung and chest areas within the 3D CT volume using linear combination of Gaussian (LCG), along with their spatial probability model, generated based on a 3D Markov Gibbs random field (MGRF), with potentials estimated analytically. Finally, the generated segmentation is further refined using a transformer model to maximize the log-likelihood of a given region. The proposed method’s efficacy is assessed by analyzing 3D chest scans of 28 patients diagnosed with varying severities of COVID-19. Four metrics are employed for evaluation, namely, Dice similarity coefficient (DSC), overlap coefficient, Hausdorff distance (HD), and absolute volume difference (AVD). The proposed system demonstrates outstanding performance, with scores of \(95.60_{\pm 1.37}\%\) , \(91.61_{\pm 2.51}\%\) , \(6.56_{\pm 2.68}\) , and \(6.23_{\pm 3.52}\) , respectively. These findings underscore the potential of the proposed system in delineating both normal and pathological lung regions in CT images, when compared to its individual components as well as six state-of-the-art (SOTA) segmentation methods.