Research Projects

Pheochromocytomas and paragangliomas (intra- and extra-adrenal tumors, respectively; PPGLs) are neuroendocrine tumours arising from chromaffin cells whose pathogenesis and progression are greatly regulated by their genetics. The genetic spectrum of PPGLs can be broadly categorized into four groups: cluster 1A (SDHx), cluster 1B (VHL/EPAS1), cluster 2 [kinase signaling (KS)], and sporadic. Identifying PPGL’s genetic cluster is essential as clinical management and outcomes vary based on genetic cluster. However genetic testing for PPGLs is currently expensive and time-consuming. CT scans are acquired at the beginning of patient management for PPGL staging and determining the next therapeutic steps. Moreover, biochemical profiling is carried out in the initial stage of the treatment. Thus, we aim to demonstrate a multi-modal deep learning-based scheme that uses contrast-enhanced CT (CE-CT) scans and biochemical profiles of PPGL patients. Recently, we collected the CE-CT scans of 285 patients, who underwent screening at NIH. These CE-CT images for four genetic clusters show multiple challenges of class imbalance, intra-class variance, inter-class similarity, and large tumor size variation. We also collected the biochemistry results (metanephrine/epinephrine/methoxytyramine) and the personal details (gender, age, ethnicity) of the respective PPGL patients to curate the multimodal dataset comprising CE-CTs and biochemistry results. The standard of reference for each tumor included its genetic cluster from genetic testing, and its anatomical location (head and neck, chest, adrenal, and abdominopelvic excluding adrenal). We intend to evaluate the proposal with accuracy, balanced accuracy, F1-score & AUC. Recently we showed that it is difficult to identify PPGL’s genetic clusters only from CE-CTs. Medical experts often look at the biochemistry profiles to get a rough idea about PPGL’s genetic category. Thus, fusing the biochemistry results with the CE-CTs could lead us to enhance the PPGL’s genetic cluster identification performance. We propose to develop a deep learning-based model which must be capable of handling both the CT image and the biochemistry profile including patients’ details as its input. We will train an attention-based network (like a transformer) for extracting features from CT images of PPGLs. Similarly, we will train an encoder-decoder network with the PPGLs’ biochemistry results and other details. We will be also exploring self-supervised learning strategies with the PPGL’s anatomical location information to guide the encoder-decoder networks by defining a suitable pre-text task. Finally, the features extracted from the networks corresponding to CT images and biochemistry profiles will be fused to train a classifier for the 4-way classification of PPGL’s genetic clusters. We will study all of the early, intermediate, and late fusion techniques to come up with a unique and efficient model for PPGLs’ genetic group identification.