A fully automatic segmentation pipeline of pulmonary lobes before and after lobectomy from computed tomography images

Abstract

Background and objective: Lobectomy is a curative treatment for localized lung cancer. The study aims to construct an automatic pipeline for segmenting pulmonary lobes before and after lobectomy from CT images.

Materials and methods: Six datasets (D1 to D6) of 865 CT scans were collected from two hospitals and public resources. Four nnU-Net-based segmentation models were trained. A lobectomy classification was proposed to automatically recognize the category of the input CT images: before lobectomy or one of five types after lobectomy. Finally, the lobe segmentation before and after lobectomy was realized by integrating the four models and lobectomy classification. The dice similarity coefficient (DSC), 95% Hausdorff distance (HD₉₅) and average symmetric surface distance (ASSD) were used to evaluate the segmentations.

Results: The pre-operative model achieved an average DSC of 0.964, 0.929, 0.934, and 0.891 in the four datasets. In D1 and D2, the average HD₉₅ was 4.18 and 7.74 mm and the average ASSD was 0.86 and 1.32 mm, respectively. The lobectomy classification achieved an accuracy of 100%. After lobectomy, an average DSC of 0.973 and 0.936, an average HD₉₅ of 2.70 and 6.92 mm, an average ASSD of 0.57 and 1.78 mm were obtained in D1 and D2, respectively. The postoperative segmentation pipeline outperformed other counterparts and training strategies.

Conclusions: The proposed pipeline can automatically segment pulmonary lobes before and after lobectomy from CT images and be applied to manage patients with lung cancer after lobectomy.

Keywords: Computed tomography; Convolutional neural network; Pulmonary lobe segmentation; Pulmonary lobectomy.