Deep learning and habitat radiomics based on contrast-enhanced CT for differentiating pathological subtypes of small renal cell carcinoma

This study aimed to develop a deep learning model, a peritumoral radiomics model, a habitat radiomics model, and a conventional clinical-imaging model, and to explore their value in the preoperative differentiation of common pathological subtypes of small renal cell carcinoma (sRCC). A total of 143 patients with pathologically confirmed sRCC (92 clear cell RCC and 51 non-clear cell RCC) were retrospectively included. Patients were randomly divided into a training set ( n = 101) and a validation set ( n = 42) at a ratio of 7:3. Corticomedullary phase images were used for analysis. Tumor regions of interest (ROIs) were manually delineated using ITK-SNAP software, and peritumoral ROIs were generated by uniform 2-mm expansion. Unsupervised K-means clustering was performed on the intratumoral ROI based on local features to identify intratumoral subregions, and the optimal number of subregions was determined using the maximum Silhouette Coefficient. Radiomics features were extracted from peritumoral region and each subregion. The maximum cross-section of the tumor was obtained. Transfer learning was then applied to initialize the ResNet18 model with pre-trained parameters from the ImageNet dataset, and deep transfer learning features were subsequently extracted. Feature selection and dimensionality reduction were conducted using the max-relevance and min-redundancy (mRMR) algorithm and the least absolute shrinkage and selection operator (LASSO) method, with cross-validation applied within the training set. Univariate and multivariate logistic regression analyses were used to identify the independent clinical predictors. Based on the identified independent clinical predictors and the selected features, clinical, peritumoral, habitat, and deep learning models were developed. The model with the highest performance was combined with independent clinical predictors to build the combined model. The performance of the models was evaluated using the area under the curve (AUC), accuracy, sensitivity, and specificity, and clinical utility was assessed using decision curve analysis. The optimal number of clusters was 3. Logistic regression analysis identified cystic or necrotic areas and enhancement characteristics as independent clinical predictors. A total of 8 peritumoral radiomics features, 12 habitat radiomics features, and 7 deep learning features were ultimately selected. The deep learning model (AUC: 0.960, 95% CI: 0.927–0.993 in the training set; AUC: 0.968, 95% CI: 0.926–1.000 in the validation set) outperformed the clinical model (AUC: 0.798, 95% CI: 0.714–0.883; AUC: 0.794, 95% CI: 0.647–0.940), the peritumoral model (AUC: 0.885, 95% CI: 0.815–0.955; AUC: 0.886, 95% CI: 0.776–0.997), and the habitat model (AUC: 0.951, 95% CI: 0.904–0.998; AUC: 0.943, 95% CI: 0.881–1.000). The performance of the combined model, established by integrating the deep learning model with the clinical model, was further improved (AUC: 0.968, 95% CI: 0.939–0.997 in the training set; AUC: 0.973, 95% CI: 0.934–1.000 in the validation set). Decision curve analysis demonstrated that the combined model provided the highest net benefit. The radiomics and deep learning models based on contrast-enhanced CT enabled effective noninvasive differentiation of sRCC subtypes. Habitat radiomics can capture intratumoral heterogeneity and shows favorable diagnostic performance, while deep learning demonstrates superior discriminative ability. These approaches may assist in preoperative decision-making, although further validation in larger external cohorts is warranted.