Ultrasound imaging is the most widely used tool in cancer screening and diagnosis. In addition to conventional B-mode ultrasound, advanced ultrasound techniques such as color Doppler and elastography imaging have become increasingly prevalent in clinical practice, substantially enhancing diagnostic accuracy. With the development of artificial intelligence, multimodal ultrasound image based deep learning has become an effective approach to provide consistent and accurate interpretation. However, due to factors such as patient non-compliance or equipment constraints, advanced ultrasound modalities may be unavailable, which can negatively affect model performance. To address this challenge, we propose Ultrasound Feature Imputation Model (USFIM), a feature imputation method based on conventional B-mode ultrasound. USFIM leverages features extracted from B-mode images to predict and impute the features of missing modalities and then incorporates knowledge distillation to further enhance the imputation quality. We validate the effectiveness of USFIM in the tasks of breast, thyroid, and liver cancer predictions. In addition, we demonstrate the generalizability of the proposed approach via a large external test cohort. In conclusion, this study contributes to advancing ultrasound image analysis while reducing dependence on costly multimodal imaging equipment in clinical practice.