Fusion of habitat analysis and deep learning on contrast-enhanced T1-weighted imaging for predicting Ki-67 status in pediatric brain tumors

Tumors are heterogeneous and consist of subregions, also known as tumor habitats, each of which corresponds to a group of tissues with similar structural, metabolic or functional characteristics. This study aims to visualize and quantify intratumoral heterogeneity using habitat and deep learning models, and to evaluate the performance of single and fusion models in predicting the Ki-67 index of pediatric brain tumors based on contrast-enhanced T1-weighted imaging (CE-T1WI) clinical parameters, thereby guiding treatment and assessing patient prognosis.

Retrospectively, CE-T1WI images from 140 pediatric patients were analyzed. Habitat analysis and Deep Learning features were extracted from the CE-T1WI data. Feature selection was performed using Pearson correlation, least absolute shrinkage and selection operator (LASSO), and principal component analysis (PCA). Random forest algorithms were utilized to construct habitat and Deep Learning models separately. Subsequently, optimal features from both models were merged to develop a fusion model for predicting Ki-67 index. The predictive performance of the habitat, Deep Learning and fusion models was evaluated using the area under the receiver operating characteristic curve (AUC).

Patients were divided into training and test cohorts at a 7:3 ratio. In the training set, the habitat and Deep Learning models achieved AUCs of 0.95 and 0.91, respectively, while in the validation set, the AUCs were 0.74 and 0.69. The fusion model demonstrated superior performance, with AUCs of 0.97 in the training set and 0.83 in the validation set for predicting the Ki-67 index.

Habitat imaging effectively quantifies intratumoral heterogeneity and provides a promising approach for the quantitative and qualitative characterization of different tumor subregions. The fusion model, integrating habitat and Deep Learning features derived from CE-T1WI, serves as an excellent tool for predicting Ki-67 index in pediatric brain tumors. This approach can assist clinicians in efficiently predicting patients' Ki-67 index, facilitating targeted decisions for subsequent treatments.