BACKGROUND: Nuclear medicine imaging allows for a wide variety of data acquisition and image generation methods in the clinical setting. Imaging phantoms are routinely used to evaluate and optimize image quality and quantitative accuracy of features, but few phantoms realistically model the anatomy or heterogeneity of target regions within patient images, such as tumours that are commonly observed in oncology. We developed a negative cast modelling (NCM) technique which enables applications such as non-standard shape tumour phantoms, organ phantoms for radiation dosimetry, and quality control phantoms with small lesions. METHODS: Tumour templates were derived from segmented PET images of primary mediastinal B-cell lymphoma (PMBCL) patients. Lesion segmentations were saved and 3D-printed. Negatives were developed using silicone-based molding materials, and final models cast using a composition of liquid plastic, pigment, and PET radiotracer. Images of lesions were acquired using the GE DMI PET/CT scanner, and image features were quantified. RESULTS: Mean absolute error (MAE) for tumour volume between the original template and casted models is 13.8%, indicating that the method is reasonably accurate. The high viscosity of the liquid plastic used in the casting process establishes non-uniform tumour models, which is very useful in practice for evaluating image features related to heterogeneity. PET images using the NCM method is determined to be highly realistic by an experienced nuclear medicine physician, due to the non-standard shapes that can be established within the tumours. CONCLUSIONS: The NCM method has potential to enable more realistic phantom studies within nuclear medicine imaging. The cost for the lymphoma tumour phantom study is less than $400 USD, making it feasible for large-scale studies.