Automating the Assessment of Patients with Lymphoma
Research Description:
Diagnosis and therapeutic response assessment in patients with lymphatic diseases using FDG positron emission tomography/computed tomography (FDG PET/CT) is difficult due to the varied appearance and location of lymph tissue throughout the body. An accurate assessment of disease is important, as treatment courses may change based on the suspicion of a single lymph node. However, current clinical standards require only a qualitative assessment of the patients, which is known to be highly subjective and does not take full advantage of the information in PET images. The overall goal of my work is to develop a comprehensive, automatic analysis method to quantify and assess patients with lymphoma.
Localization of diseased lymph nodes on FDG PET images is complicated by high levels of physiologic uptake throughout the body. We are thus using deep learning methods to automate the process of disease detection, as shown in Figure 1. Working with lymphoma specialists from several institutions, we have acquired a large database of lymphoma patients to train and test the model, comparing to a physician-based ground truth. Once localized, disease is automatically segmented, and features quantifying disease burden and spatial heterogeneity can be extracted and tracked over time to monitor patients and predict response.
Figure 1: The convolutional neural network (DeepMedic) is trained to automatically locate disease lymph nodes.
Cancer patients may be monitored with PET imaging for long periods of time, often on PET scanners with varying quantitative abilities. Thus, another aspect of my work is to ensure quantitative consistency of patient images throughout their monitoring period. To do this, we perform what is called PET scanner harmonization, where PET images acquired on newer scanners are reconstructed with settings shown to reduce quantitative differences in a phantom. There is a wide range of complexity levels for PET imaging phantoms, as shown in Figure 2. Our group aims to quantify the impact that scanner harmonization can have on patient response assessment, as well as simplify the process of harmonization so it is more accessible to a larger number of clinics.
