Clustering Model Based on 3DUNET-Gmm Combining 3D-Unet Feature Extracting with Gmm Clustering to Divide Tumor Subregions on FDG PET for La-NSCLC
Abstract
Purpose
Accurate identification of high-risk and low-risk tumor subregions enables radiographers to customize radiation dose distributions for biologically adaptive therapies. This study proposes a 3DUNET-GMM model that integrates 3D-UNet feature extracting with Gaussian mixture model (GMM) clustering to divide tumor subregions on FDG PET for locally advanced non-small cell lung cancer (LA-NSCLC).
Methods
25 LA-NSCLC patients enrolled on the FLARE-RT trial (NCT02773238) underwent FDG-PET/CT imaging prior to (PrePET) and during week 3 (MidPET) of chemoradiotherapy. PET voxel grids were mapped into a fixed 1283 grid through bounding-box isotropic scaling of spatial coordinates. The 3DUNET-GMM model extracted voxel-level feature maps (standardized uptake value (SUV), coordinates) by 3D-UNet feature extracting, and used GMM clustering to cluster the feature points on feature maps into 2 clusters. Clustering performance was evaluated using the Calinski-Harabasz index (CH), silhouette coefficient (SC) and Davies-Bouldin index (DB) under leave-one-patient-out cross-validation. Considering that the risk level division depends on the average SUV within the 2 clusters, clustering metrics were calculated under two distance-definition schemes. Scheme 1 calculated the distance between two feature points using SUV only. Scheme 2 incorporated both SUV and spatial coordinates to calculate the distance.
Results
In scheme 1, the clustering model based on 3DUNET-GMM achieved superior CH, SC and DB (CH_pre=3343436.88, SC_pre=1.00 and DB_pre=0.07; CH_mid=4349205.32, SC_mid=1.00 and DB_mid=0.09) compared with GMM (CH_pre=2138.60, SC_pre=0.13 and DB_pre=8.25; CH_mid=6800.19, SC_mid=0.24 and DB_mid=3.44). In scheme 2, the 3DUNET-GMM model demonstrated higher SC and lower DB (SC_pre=0.54 and DB_pre=0.93; SC_mid=0.54 and DB_mid=0.95) than GMM (SC_pre=0.22 and DB_pre=2.27; SC_mid=0.21 and DB_mid=2.29). The clustering performance of 3DUNET-GMM was consistently numerically better than performance achieved by conventional GMM clustering model.
Conclusion
The 3DUNET-GMM framework enables effectively delineation of high-risk and low-risk tumor subregions on FDG PET and outperforms the conventional GMM clustering model. This approach may support biologically-adaptive radiotherapy and enhance treatment assessment.