Voxel-Level Segmentation of Rois In Cervical Cancer Patients Based on Reproducible Radiomics Features
Abstract
Purpose
Radiotherapy is one of the primary treatment modalities for cervical cancer, and accurate ROI delineation is crucial. Previous studies have predominantly utilized feature reproducibility to construct patient prognosis models, with limited exploration of voxel-level Habitat segmentation. This study aims to evaluate feature reproducibility in clinical settings and achieve stable voxel-level Habitat segmentation.
Methods
We retrospectively included 356 planning CT scans from two hospitals, with ROIs manually delineated by clinicians. Using image perturbation strategies simulating patient position variations and delineation differences, we extracted 1,302 radiomics features from raw, 5 Log-filtered, and 8 wavelet-filtered images. Features were analyzed across two dimensions: kernel radius (1, 3, 5 mm) and bin width (8, 16, 32, 64, 128 HU). Redundant features with Spearman correlations > 0.9 were excluded. Intraclass correlation coefficients (ICC) were calculated using 1000 bootstrap resamples, with the lower confidence limit (ICC_LCL) > 0.75. ROI-specific repeatable features were selected to generate voxel-level feature maps. Gaussian mixture models (GMM) were used to cluster Habitats, with Dice similarity evaluating segmentation spatial consistency between original and perturbed images.
Results
Feature repeatability peaked at kernel = 1mm and Bin Width = 32HU, yielding ICC_LCL (Median[IQR]: 0.966[0.947-0.982]). High stability was achieved using LoG filtering (3-5mm) and wavelet filtering (LLL, LHL, LLH, LHH) in the image generation domain. Ultimately, 19, 14, and 20 features were selected for Rectum, Bladder, and CTV, respectively. The Repeatability group constructed habitats (Habitat1, Habitat2, Habitat3) showed median [IQR] Dice coefficients of 0.874 [0.595–0.940], 0.921 [0.696–0.974], and 0.913 [0.786–0.977], respectively, significantly higher than the All group (0.757 [0.525–0.897], 0.885 [0.756-0.969], 0.897 [0.710-0.967]), with statistically significant differences (P < 0.001).
Conclusion
Incorporating repeatability features significantly enhances the stability of voxel-level Habitat segmentation, offering the potential to establish more reliable markers for assessing imaging heterogeneity and providing potential value for clinical practice applications.