Purpose To evaluate the predictive utility of radiomic features extracted from ultrashort echo time (UTE) MRI in comparison to conventional proton density (PD) sequences for short-term (24-week) and long-term (5-year) clinical outcomes in patients with patellar tendinopathy (PT) receiving exercise therapy. Materials and methods This prospective study of 76 PT patients undergoing 24-week exercise therapy underwent baseline 3D UTE and PD MRI at 3.0 T. The patellar tendon segmentation used nnU-Net, evaluated with Dice coefficient. Six predictive models consisting of clinical covariates and radiomic features from UTE and PD were developed using Elastic Net with 10-fold cross-validation. Model performance in predicting responsiveness of the patient-reported Victorian Institute of Sports Assessment (VISA-P) score was evaluated using the area under the receiver operating characteristic curve (ROC AUC) and the precision-recall curve (PR AUC), with 95% confidence intervals. Results The mean Dice similarity coefficient for the automatic segmentation of the patellar tendon from 3D-PD was 0.92 (SD: 0.02) and from 3D-UTE-Cones 0.89 (SD: 0.03). The UTE-based radiomics model demonstrated the highest predictive performance at 24 weeks (ROC AUC: 0.714 [95% CI: 0.701–0.727]; PR AUC: 0.848 [0.837–0.858]), while the PD-based model showed the lowest (ROC AUC: 0.569 [0.553–0.584]; PR AUC: 0.710 [0.692–0.727]). At the 5-year follow-up, UTE radiomics maintained robust performance (ROC AUC: 0.692 [0.677–0.706]; PR AUC: 0.822 [0.810–0.834]), whereas PD radiomics remained limited (ROC AUC: 0.578 [0.561–0.594]; PR AUC: 0.694 [0.676–0.713]). Conclusions Radiomics features extracted from UTE MRI demonstrate the highest predictive performance for clinical outcomes.