Planning and model-based control of noninvasive thermal therapies require site-and patient-specific thermal response and power deposition models of the target. In this paper, we describe an approach to the identification of low-dimensional treatment models based on magnetic resonance temperature imaging of the patient. First, proper orthogonal decomposition (POD) of MRI thermal images is used to identify a reduced set of basis functions, which capture spatial correlations within images. Next, the multivariate autoregression of image projections into the manifold, spanned by reduced-order POD basis, is used to identify low-dimensional discrete models of patient's thermal response and specific absorption rate (SAR) of noninvasively transduced energy. Compared to the results of [1], [2], the developed approach is more suitable for real time model reidentification, which may be necessary during high-temperature therapies, such as high-intensity focused ultrasound (HIFU) and other ablation treatments, which are known to substantially change blood perfusion and tissue properties. The developed approach was validated during MR thermal imaging experiments with tissue phantom noninvasively heated by focused ultrasound. The results demonstrate that the developed method utilizes medical imaging to identify accurate low-dimensional models of therapies suitable for treatment planning and modelbased treatment control with imaging feedback.