Low-back musculoskeletal disorders are commonly associated with elevated muscle tone and region-dependent variations in soft-tissue stiffness, which pose significant challenges for safe and effective robot-assisted therapy. To support clinicians in delivering consistent and compliant treatment, this paper presents a vision-assisted, region-aware stiffness adaptation framework for robot-assisted lumbar therapy An RGB-D camera mounted on the robot end-effector is used to identify discrete anatomical regions through visual markers, enabling reliable localization of muscular, boundary, and bony areas without external tracking systems. Based on the detected region label and real-time force feedback, a quadratic-programming (QP) layer adaptively adjusts the normal-direction stiffness of a Cartesian impedance controller to achieve appropriate compliance across different tissue regions. To ensure interaction safety during stiffness modulation, an energy-tank mechanism is incorporated to enforce passivity by constraining stiffness updates according to the available interaction energy. The proposed approach is validated on a biomechanically realistic lumbar phantom using a collaborative robot platform. Experimental results demonstrate that region-aware stiffness adaptation improves force regulation compared to fixed-stiffness impedance control, while the energy-tank mechanism effectively suppresses abrupt stiffness changes and maintains energetically consistent behavior during contact-rich interactions. These results suggest that the proposed framework provides a practical and safety-oriented approach for vision-guided, compliant medical robot interaction.