Accurate segmentation of cancerous lesions remains a major challenge in medical imaging, particularly in scenarios involving multifocal, small, and variable-sized lesions. While state-of-the-art deep learning methods have demonstrated strong performance in organ-level or large-lesion segmentation, their effectiveness diminishes in clinically critical tasks such as identifying lymphoma in PET/CT or brain metastases in MRI, where lesions are numerous, heterogeneous, and often subtle. In this work, we investigate the development of a foundational model specifically tailored to the segmentation of cancerous lesions under these challenging conditions. Our approach integrates both automatic and interactive strategies: the model is designed to perform robust, fully automated lesion segmentation when feasible, but also supports minimal user interaction to refine and guide results in difficult cases. This hybrid paradigm reflects the emerging direction of the field, acknowledging that complex oncological imaging tasks may resist complete automation while still benefiting from the efficiency of deep learning–driven methods. By targeting small and multifocal cancerous lesions, this research aims to bridge the current performance gap, offering a scalable and adaptable framework for precision oncology applications.