Diagnosing thyroid cancer is notably challenging because of its diverse manifestations and the rising number of cases worldwide. Early detection and diagnosis of thyroid nodules’ malignancy is crucial for reducing their progression. This paper introduces a novel computer-aided diagnosis (CAD) system that utilizes T2 and diffusion-weighted (DWI) magnetic resonance imaging (MRI) modalities to help diagnose thyroid cancer. First, the thyroid nodules are delineated from T2 and DWI modalities. Then, various features are extracted from these nodules, such as first order statistics (FOS), gray level co-occurrence matrix (GLCM), and gray level run length matrix (GLRLM), to capture texture and spatial information. To improve both the performance and interpretability of the model, outlier detection methods, such as the cluster-based local outlier factor (CBLOF), are utilized to identify deviations in the data. Finally, the extracted features from T2 and DWI modalities are fed into a multilayer perceptron (MLP) and LightGBM (LGBM) classifiers, respectively. Subsequently, the classifiers’ outputs are integrated using a majority fusion approach for final diagnosis. The proposed system is evaluated on 55 thyroid nodule patients using a 10-folds cross-validation approach, achieving an accuracy of \(99.48\%\) . The reported results, based on integrating decisions from each MRI modality using a majority fusion approach, clearly demonstrate the effectiveness of the proposed framework compared to the performance of well-known pretrained convolutional neural networks (CNNs).