Differentiation of cervical cancer subtypes using machine learning models on MRI images

To develop and compare various radiomics-based machine learning (ML) models for distinguishing cervical squamous cell carcinoma (SCC) from non-SCC histopathological subtypes, using multiparametric magnetic resonance imaging (MRI) and clinical data.

This retrospective study included 88 women (mean age, 51.1 ±13.0 years; range, 25-83 years) with histopathologically confirmed cervical cancer (47 SCC, 41 non-SCC). For each patient, axial and sagittal T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) were available, along with clinical metadata. Radiomic features (shape, first-order, and texture features including gray-level co-occurrence matrix, gray-level run length matrix, and gray-level size zone matrix) were extracted from each sequence, yielding 945 features per patient. A feature set was created by combining features from all three sequences. The least absolute shrinkage and selection operator (LASSO) method was used for feature selection. Three supervised classifiers - random forest (RF), support vector machine (SVM), and logistic regression (LR) - were trained to classify SCC versus non-SCC.

Among the single-sequence models, the sagittal T2WI sequence demonstrated the strongest performance, achieving receiver operating characteristic (ROC) area under the curve (AUC) values of 0.839 with SVM and 0.732 with LR. In contrast, the axial T2WI RF model showed poor discriminative ability (ROC-AUC: 0.464), indicating that this sequence alone provides limited value for subtype differentiation despite showing acceptable accuracy metrics. The combined multi-sequence model yielded the highest overall performance. Using LASSO-selected features, RF achieved the best ROC-AUC (0.958), followed closely by SVM (0.953) and LR (0.951), with all models attaining accuracies above 0.91.

Feature-level integration of axial T2WI, sagittal T2WI, and DWI substantially enhances ML performance in differentiating SCC from non-SCC cervical cancer subtypes. Compared with single-sequence models, the combined multi-sequence approach provides more robust and well-balanced classification, underscoring the complementary value of multiparametric MRI for histopathological subtype prediction.