Related Experiment Video
Updated: Jun 13, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Comparing Habitat, Radiomics, and Fusion Models for Predicting Micropapillary/Solid Components in Stage I Lung
Shaoyu Huang1, Xiuzhen Liang1, Kaihua Lou2
1Department of Radiology, The Affiliated Lihuili Hospital of Ningbo University, 315010 Ningbo, Zhejiang, PR China (S.H., X.L., J.Z., J.W., H.D.).
Rationale And Objectives:
To comprehensively compare habitat, radiomics, and fusion models for the preoperative prediction of micropapillary/solid (MP/S) status in stage I lung adenocarcinoma (LAC).
Materials And Methods:
In this retrospective study, we enrolled 345 patients postoperatively diagnosed with stage I LAC from two medical centers, dividing them into training (n=207), internal validation (n=69), and external validation (n=69) cohorts. Radiomics model (RM) was developed using CT images of the primary tumor. Habitat model (HM) was built by analyzing intra-tumor subregions identified via unsupervised K-means clustering algorithm. Fusion model employed two integration strategies as follows: feature-based pre-fusion model (pre-FM) and decision-based post-fusion model (post-FM). The predictive performance of all models was comprehensively evaluated by area under the curve (AUC) and integrated discrimination improvement (IDI). Additionally, correlations between clustering and radiomics features were analyzed with Spearman's correlation analysis.
Results:
The HM demonstrated superior predictive performance compared to the RM in the training cohort (AUC: 0.900 vs. 0.876, p=0.252). The pre-FM consistently outperformed the HM and RM across all study cohorts (AUC: 0.843-0.914 vs. 0.802-0.900 and 0.841-0.876, p=0.041-0.484 and 0.011-0.924, respectively). The post-FM further enhanced predictive performance, as evidenced by the highest AUCs in the training and internal validation cohorts (AUC: 0.952 vs. 0.862-0.914, p=0.001-0.116; 0.850 [0.724-0.922] vs. 0.770-0.843, p=0.102-0.922). and IDI values (14.2%-36.4% increase). Additionally, clustering and radiomics features displayed a higher number of correlated feature pairs in the MP/S (+) group than MP/S (-) group.
Conclusion:
The post-FM, integrating clustering signature, radiomics signature, and clinical characteristics, has been established as a reliable predictor for MP/S status in stage I LAC.
