Accuracy assessment of a hybrid deep learning and image processing approach for lung CT segmentation in external radiotherapy

Abstract

Background : Accurate lung segmentation in computed tomography (CT) is a fundamental step for numerous clinical applications, including radiotherapy planning and thoracic disease diagnosis. Manual segmentation, however, remains time-consuming and operator-dependent. This study aims to develop a fully automated segmentation framework combining a well-structured pipeline of image processing techniques with a deep learning model based on a U-Net architecture and a pre-trained VGG16 encoder. Materials and methods : Ninety anonymized thoracic CT scans, including lung and breast cancer patients, were collected, pre-processed and used to generate binary lung masks using a custom Python pipeline. The dataset was split patient-wise into training (70%), validation (15%) and testing (15%) subsets. The model was trained using binary crossentropy loss, the Adam optimizer and data augmentation strategies to improve generalization. Results : High correlations were observed between automatically computed lung volumes and mean hounsfield units (HU) compared to the treatment planning system (TPS) references (R 2 = 0.84 for volume, R 2 = 0.72 for HU, both with p < 1 × 10 –33 ) and the deep learning model achieved a mean dice similarity coefficient (DSC) of ~95% and intersection over union (IoU) of ~95% on an independent test set. Mean loss and pixel-wise accuracy were 0.004 and ~95%, respectively. Conclusions : The proposed VGG16-based U-Net model demonstrates high accuracy and robustness in segmenting lungs from thoracic CT scans, achieving ~95% DSC and ~99% pixel-wise accuracy. It effectively replaces manual segmentation with a fully automated, scalable solution. This framework streamlines lung segmentation in clinical radiotherapy workflows, reducing operator variability and time consumption.