Quantitative evaluation of spatially-variant deformations recovered by deep learning on clinical-like breast lesions

The dedicated dual-panel breast PET scanner (B-PET) can potentially provide improved imaging of breast lesions with higher spatial resolution and sensitivity and lower scanner costs. On the other hand, the dual-panel systems suffer from spatially-variant deformations, due to the limited angular sampling and parallax errors, that need to be mitigated in order to improve visualization and quantification of the breast lesions. Our previous studies demonstrated that the deep-learning approach can significantly reduce spatially-variant deformations in the context of B-PET. However, these studies used only simplified geometric objects with a uniform background for training and testing, which does not fully represent the clinical complexity in breast PET imaging. In this work, we developed a methodology for generation of synthetic clinical-like breast images with complex lesion shapes, including spiculated lesions and following tracer-dependent activity characteristics for [¹⁸F]FES and [¹⁸F]FDG. Subsequently, the neural network was trained and tested on clinical-like synthetic B-PET data reconstructed using statistical iterative reconstruction (DIRECT-RAMLA) used as the network input. Our results show that the deep learning approach can substantially suppress deformations for both small lesions and the larger phantom itself in dual-panel PET reconstructions. The deep learning approach can also improve quantitative measurements in terms of the lesion contrast metrics and image roughness over the phantom background. A research [¹⁸F]FES patient study further confirms the improved visual image quality and a good visual correlation for a heterogeneous lesion with the total body PET imaging result.