A grayscale ultrasound-based two-stage deep learning framework for automatic segmentation and benign-malignant differentiation of subpleural pulmonary lesions

Abstract

Lung cancer is one of the most lethal malignancies globally. However, accurately differentiating subpleural malignant lesions from benign infectious conditions via grayscale ultrasound remains difficult due to operator subjectivity. We developed and validated a two-stage deep learning framework for automated lesion segmentation and benign-malignant classification. In this retrospective study, the research data were derived from patients with SPLs who attended Center 1 between Jan 2023–Dec 2024, as well as patients recruited from Center 2 and Center 3 Jun 2024–Dec 2024 (serving as External Test Set 1 and 2, respectively). The nnU-Net, U-Net and DeepLabv3 + models were adopted for lesion segmentation, while the DenseNet121/201, EfficientNet-B0/B1/B2/B3/B4/B5, ResNet18/50/101, Inception-V3 and VGG19 models were used for benign-malignant classification. Model performance was evaluated using metrics including the Dice similarity coefficient and area under the receiver operating characteristic curve (AUC). Diagnostic performance of 6 radiologists (3 junior, 3 senior) with varying experience was compared between unaided and AI-assisted readings; pooled three-reader differences were estimated by case-cluster bootstrap with Bonferroni adjustment. A total of 1059 patients were included in the study (61.32 ± 13.82 years; 726 male patients). The nnU-Net model achieved optimal segmentation (Dice: 0.920 [tuning-validation set, n  = 257], 0.913 [External Test Set 1, n  = 108], 0.915 [External Test Set 2, n  = 94]). Among thirteen candidate classification architectures, DenseNet121 yielded the best performance on the internal tuning-validation set and was therefore selected as the final model; it was then evaluated on two external cohorts, achieving AUCs of 0.889 (95% CI: 0.827–0.951) and 0.865 (95% CI: 0.790–0.940) in External Test Sets 1 and 2, respectively. Overall diagnostic accuracy was significantly higher in the AI-assisted reading session for junior radiologists in both external test cohorts (both Bonferroni- adjusted P  < 0.001), whereas improvements among senior radiologists did not reach statistical significance after multiplicity adjustment. The deep learning model constructed in this study can accurately realize automatic segmentation and benign-malignant classification of SPL ultrasound images, serving as a potential auxiliary diagnostic tool. Not applicable.