Deep-learning based quantitative evaluation of postoperative atelectasis following right upper lobectomy

Existing methods of grading atelectasis are typically subjective and not scalable. We aimed to develop an automated, deep learning-based framework to quantify and grade postoperative atelectasis. We retrospectively included all patients who underwent RULobectomy from 2008 to 2023. We trained three nnU-Net v2 segmentation models for preoperative and postoperative lobes and airways with volumetric quantification of the right middle lobe (RML), right lower lobe (RLL), and total lung volume. Atelectasis severity in the RML was independently graded using a 5-point radiological scale (none, minimal, subsegmental, segmental, lobar). The association between volume metrics with atelectasis severity and clinical outcomes was evaluated. 236 patients comprised the study cohort. Median(IQR) RML volume loss progressively increased with higher atelectasis grades, from -4.6 mL (-78.5, 59.0) in grade 0 to -317.8 mL (-440.7, -194.8) in grade 4 atelectasis (p < 0.001). Normalized RML/right lung (RL) and RML/total lung (TL) volume ratios showed statistically significant differences across the pooled atelectasis grades (p < 0.001). Normalized RLL volumes increased with worsening RML atelectasis (p < 0.001), suggesting compensatory hyperinflation. A higher ΔRML/RL [OR(95%CI): 0.89 (0.81-0.98), p = 0.01] and ΔRML/TL [0.80 (0.65-0.98), p = 0.03] were associated with reduced 1-year need for bronchoscopy. We demonstrate the feasibility and clinical relevance of deep learning-based volumetric assessment of atelectasis after RULobectomy.