PROJECT SUMMARY 18F-Fluorodeoxyglucose (FDG) PET/CT imaging has become an essential tool for guiding and adapting treatments for lymphoma. However, the PET evaluation criteria currently used for assessing lymphoma, which consists of subjective visual scoring on a 5-point scale, is suboptimal. The visual scores suffer from high inter- observer variability and have low prognostic power for new emerging biological therapies. Quantitative PET metrics have been shown to be more predictive of clinical outcomes than visual scores, but quantitative analysis of whole-body PET/CT images is prohibitively time-consuming and impractical in routine clinical care. Deep learning (DL) has shown promise in automating the quantitative analysis of baseline FDG PET/CT images, but comprehensive evaluation of interim-therapy and post-therapy images using DL has proven difficult. Residual lymphoma has low-level uptake, which can be hard to differentiate from physiologic or treatment-related uptake, and reading physicians must use clinical histories and baseline PET images (i.e., sites of initial disease) to make reliable diagnoses. DL algorithms, on the other hand, only operate on cross-sectional images and are unable to account for historical context. Our objective is to develop DL algorithms that operate on PET/CT images from more than one time point so that algorithms can learn longitudinal dependencies for contextually-aware predictions. We also aim to develop multimodal vision-language models that can simultaneously interpret radiology text reports while performing PET/CT image analysis. These models can leverage critical information about patient history and physician interpretation when processing retrospective images. Furthermore, we will use semi-supervised learning to leverage both unlabeled datasets and labeled datasets. Our overall goal is to develop contextually-aware algorithms for automated longitudinal analysis of whole-body PET/CT images in lymphoma. These tools will be developed using diverse datasets from multiple institutions. PET metrics measured by DL will be validated as predictive markers of outcome using data from a Phase 3 clinical trial.

PROJECT NARRATIVE This proposal will improve the automatic analysis of longitudinal, multi-time point PET/CT images of lymphoma by developing contextually-aware and multimodal vision-language deep learning algorithms. Quantitative PET metrics are highly predictive of outcome for lymphoma and could be used to improve the efficacy of PET-adapted therapy, but quantification of whole-body PET/CT images is time consuming and automated analysis of interim- therapy images is technically challenging. This project will address these challenges by designing and validating deep learning PET analysis algorithms that can simultaneously consider baseline and interim PET scans, thus mimicking the processes used by physicians, and also by designing semi-supervised multimodal algorithms that can use the information in radiology narrative reports for enhanced image analysis.