Development of a prototype software for automated PET/CT interpretation and reporting in thoracic cancer

点击翻译按钮获取中文摘要

Abstract. In cancer, body-wide FDG-PET/CT is a prime modality for diagnosis, staging, and treatment assessment. Despite its paramount importance to enable precision medicine in cancer, no method is currently available for automated disease burden estimation and standardized reporting on PET/CT images regionally and globally in anatomic organs and lymph node zones within a body region or body-wide. Automated production- mode body-wide/ body-region-wide disease measurement with standardized reporting will foster cancer research discovery and will be of great interest to oncologists, radiologists/ nuclear medicine physicians, Medicare and private health insurers, and pharmaceutical companies that conduct clinical trials of new cancer therapeutics and currently rely on manual methods of response assessment. The overarching goal of this Phase I project is, therefore, to develop, validate, and demonstrate a prototype software for disease measurement and reporting via FDG-PET/CT in the above manner in one body region, namely thorax, based on innovative algorithms that are generalizable body-wide. The project has two aims: Aim 1: Develop, implement, and validate algorithms for disease burden estimation in thoracic cancer via FDG-PET/CT. Aim 2: Develop and demonstrate a prototype software implementing the above algorithms for disease measurement and reporting. Aim 1 will be accomplished in 3 stages: Tasks 1, 2: PET/CT image data sets which are radiologically near normal for the thoracic body region will be gathered from existing whole-body scans of 100 patients. In these data sets, 7 key anatomic organs and 5 key lymph node zones in the thorax will be delineated under expert guidance. These data will be used to build population fuzzy anatomy models following our established Automatic Anatomy Recognition (AAR) methodology. An additional 100 whole-body PET/CT scans of patients with different types of cancer will be gathered to test our methods. Using available commercial clinical software, the PET uptake properties of lesions in organs and diseased lymph nodes in lymph node zones will be measured manually and used as reference ground truth of disease burden. Task 3: Deep learning (DL) algorithms anatomically guided by AAR will be developed to very accurately localize (but not delineate) organs and lymph node zones in PET/CT images using the models. Task 4: Novel methods based on fuzzy principles will be developed to automatically tag and quantify pathological regions (without explicitly delineating them) within located organs and nodal zones, and the accuracy of disease measurement will be evaluated (Task 5). Aim 2 will be accomplished by incorporating the disease measurement methodology into a prototype software named AAR-DQ (Tasks 6, 7) based on our earlier software platform CAVASS. AAR-DQ will report disease burden in a hierarchical manner – (i) at the body-region level; (ii) at each organ/ lymph node zone level; (ii) at each lesion/ lymph node level. Expected milestones. Aim 1: AAR-DQ disease measurement not to deviate more than 10% from clinical ground truth measurement. Aim 2: Disease measurement/ reporting in under 5 minutes per patient PET/CT study.