Related Experiment Video
Updated: Sep 14, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
An interpretable biparametric MRI habitat radiomics model for predicting bone metastasis in prostate cancer: a
Juntao Gong1, Feixiang Li1, Yun Sun1
1Gansu University of Chinese Medicine, Lanzhou, China.
Background:
Bone metastasis (BM) critically determines prognosis in prostate cancer (PCa), but its noninvasive prediction remains challenging. Habitat imaging may better capture intratumoral heterogeneity and improve BM risk assessment. This study aimed to develop and validate an interpretable habitat radiomics model based on biparametric magnetic resonance imaging (bp-MRI) for predicting BM in PCa.
Methods:
This retrospective dual-center study included 238 PCa patients who underwent preoperative bp-MRI [3.0T, T2‑weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps]. Patients from the primary center were split into training (n=133) and internal test (n=57) sets; 48 patients from a second center formed an external validation cohort. Tumors were segmented into three habitat subregions via K-means clustering on ADC maps. Radiomic features were extracted from T2WI for each subregion and the whole tumor. Four models (clinical, conventional radiomics, habitat radiomics, and combined) were built using a Gaussian process (GP) classifier. Performance was evaluated using area under the receiver operating characteristic curve (AUC), DeLong test, net reclassification improvement (NRI), and calibration curves. SHapley Additive exPlanations (SHAP) analysis assessed interpretability.
Results:
The combined model achieved the highest AUCs: 0.927 (training), 0.841 (internal test), and 0.821 (external validation). The habitat model achieved AUCs of 0.870, 0.838, and 0.760 in the training, internal test, and external validation sets, respectively, outperforming the conventional and clinical models in all three cohorts. In the training and internal test sets, the DeLong test showed that the habitat and combined models had significantly better discrimination than the conventional and clinical models (P<0.05); in the external validation set, the combined model remained significantly superior, whereas the habitat model's superiority did not reach statistical significance. NRI demonstrated improved risk stratification by the habitat model. SHAP identified key predictive features from habitat subregion 2.
Conclusions:
An interpretable habitat radiomics model based on bp-MRI shows potential for predicting BM risk in PCa. The combined model provides favorable performance for clinical decision support. External generalizability requires confirmation in larger, multicenter cohorts.
