Assessment of algorithmic performance in PET-CT segmentation of solitary lung tumors

Background PET-CT (Positron emission tomography–computed tomography), a nuclear medicine imaging technique is preferably used for lung tumor detection. However, manual image analysis may be time-consuming. To enhance efficiency, many automated segmentation algorithms have been developed. This study aims to develop a radiomics agreement-based method to evaluate segmentation algorithms and to compare the performance of two deep learning-based algorithms nnU-Net (No-new-net Convolutional Neural Network) and RECOMIA (Research Consortium for Medical Image Analysis Lung Cancer “Lung Cancer FDG-PET/CT v1.5”) for PET-CT solitary lung tumor segmentation. Methods and results PET-CT images of 221 patients at the University of Debrecen’s Hospital were collected and anonymized. The predictions of segmented lung tumors done by nnU-Net and RECOMIA were compared with the manual segmented masks via Wilcoxon signed-rank test and Pearson correlation values, based on eight groups of radiomics parameters calculated by PyRadiomics software. The combined p-values were calculated using Brown’s method, which were 0.481 and < 0.001 for nnU-Net and RECOMIA, respectively. The absolute values of correlation were summed, which were 37.259 and 16.358 for nnU-Net and RECOMIA, respectively. Altogether, RECOMIA shows statistically greater differences, therefore, it performs worse than nnU-Net does. Statistical analyses with Intraclass correlation coefficients (ICC), Multivariate analysis of variance (MANOVA) and follow up pairwise t-test further confirmed stronger agreement between extracted radiomics features from nnU-Net predicted masks and manual masks than RECOMIA and manual masks, which validated the proposed comparison method. Conclusion The superior performance of nnU-Net highlights the importance of training data alignment and model adaptability in achieving accurate and reliable segmentation in medical imaging. However, nnU-Net requirements for training and extra computational demands may limit its application in some clinical environments. Despite lower accuracy, RECOMIA remains valuable because of its efficiency and intuitiveness. These findings provide valuable guidance for selecting segmentation algorithms in real-world clinical and research settings.