Over 1.6 million patients in the U.S. annually undergo chemo- or radiation- as first-line cancer therapy. After therapy, the most significant challenge for oncologists is identifying non-responders (those with residual or progressive disease), which could allow them to be switched to alternative therapies. Similarly, if those with stable or regressing disease were identified early and reliably, patients could avoid unnecessary and highly morbid surgeries or biopsies for disease confirmation. Unfortunately, expert assessment of post-treatment imaging is challenging, as residual disease is visually confounded with benign treatment-induced changes on imaging. There is hence a critical need for dedicated radiomic (computerized feature extraction from imaging) and informatics approaches to enable reliable post-treatment tumor assessment. Such tools will need to account for: (1) Limited well-curated data resources with deeply annotated pathology-validated radiographic datasets, for discovery and validation of new imaging and radiomic markers for post-treatment characterization in vivo; (2) Need for specialized radiomics tools that specifically quantify morphological perturbations in response to shrinkage/growth of the lesion for identifying progressive disease (versus benign confounders), despite presence of treatment-induced artifacts (exacerbated noise, reduced contrast, poor resolution); and (3) Lack of comprehensive quality control (QC) tools to identify which of a plethora of radiomic features are both discriminable as well as generalizable to variations between sites and scanners. To address these challenges, we propose RadxTools, a new image informatics toolkit comprising three modules: (a) RadQC to enable quality control of radiomics features across multi-site imaging cohorts, (b) RadTx comprising new radiomics tools which capture local surface morphometric changes and subtle structural deformations unique to tumor response on post-treatment imaging, and (c) RadPathFuse for creating deeply annotated learning sets by spatially mapping post-treatment changes from ex vivo surgically excised histopathology specimens onto pre-operative in vivo imaging. RadxTools will be evaluated in the context of post-treatment characterization for use cases in distinguishing (a) radiation effects from cancer recurrence for brain tumors; and (b) complete/partial vs incomplete chemoradiation response for rectal cancers. Deliverables and Dissemination: Our team has had a successful history of disseminating informatics tools (>1000 downloads), including our most recent release of RadTx which has been integrated into 3 informatics platforms. By organizing community resources and targeted workshops, as well as releasing highly curated data cohorts, our team is uniquely positioned to disseminate RadxTools to the radiomics/imaging community, professional societies, and oncology working groups. Our deliverables will include tool prototypes as modules within 5 QIN/ITCR-funded platforms (3D Slicer, MeVisLab, Sedeen, CapTk, QIFP) for widespread dissemination to targeted end-user communities, in addition to deeply annotated learning sets assembled through the 2 use-cases in this project.
This project will result in development of RadxTools, a new image informatics toolkit, with specialized modules for (a) quality control and multi-site evaluation of radiomics features, (b) creation of spatially aligned radiology-pathology datasets, and (c) quantitative characterization of treatment response on routine post- treatment imaging. These modules will enable identification and benchmarking of robust radiomic markers that accurately capture subtle phenotypes of recurrent tumor presence, thus enabling personalized follow-up in the post-treatment setting; while also accounting for multi-site variations and treatment artifacts.