Comparative analysis of U-Mamba and no new U-Net for the detection and segmentation of esophageal cancer in contrast-enhanced computed tomography images.

Abstract

Background: Radiomics research in esophageal cancer (EC) has made considerable advancements. However, manual segmentation, which is relied upon in clinical and scientific workflows, remains time-consuming and inconsistent. This study aimed to develop and validate a deep learning (DL) model for the automatic detection and segmentation of EC lesions in contrast-enhanced computed tomography (CT) images.

Methods: We retrospectively collected the CT data of patients with EC confirmed by pathology from January 2017 to September 2021 at three hospitals and from individuals with a healthy esophagus. Manual labeling of EC lesions was conducted, and DL networks [no new U-Net (nnU-Net) and U-Mamba] were trained for automatic segmentation. An optimal threshold volume for EC lesion detection was determined and integrated into the postprocessing module. The performance of DL models was evaluated in internal, external, and thin-slice image test cohorts and compared with diagnoses by radiologists. The sensitivity, specificity, accuracy, Dice similarity coefficient (DSC), and Hausdorff distance (HD) were calculated.

Results: A total of 871 patients (564 males) were included, with a median age of 67 years. DL models exhibited no significant difference from radiologists' diagnoses (P>0.05). Median DSC values for the internal, external, and thin-slice cohorts were 0.795, 0.811, and 0.797, respectively, with a corresponding HD of 9.733 mm, 7.860 mm, and 8.168 mm. An intraclass correlation coefficient greater than 0.7 was observed for 97.2% of the radiomic features extracted from thin-slice images.

Conclusions: The DL methods demonstrated exceptional sensitivity and robustness in EC detection and segmentation on contrast-enhanced CT images, not only reducing missed EC diagnoses but also providing radiologists with consistent lesion annotations.