A fully automated tool for more accurate malignancy assessment in pulmonary nodules from chest CT

PROJECT SUMMARY Each year in the United States, chest computed tomography (CT) identifies pulmonary nodules in about 1.6 million patients. Although >95% of these nodules are benign, a small fraction represent early-stage lung cancer, the early detection of which is correlated with improved outcomes. Unfortunately, conclusive assessment of ambiguous nodules often requires one or more invasive biopsies, which suffer from significant complication rates and mortality. More specific, yet equally sensitive, non-invasive follow-up methods would avoid thousands of complications or deaths each year. Computational tools using artificial intelligence (AI) can greatly improve the diagnostic performance of CT by extracting subtle information that is missed even by expert radiologists. Despite a proliferation of such AI tools with relatively high performance, they have struggled to transition into the clinic because they have been implemented in ways that increase time burden for clinicians and non-reimbursable costs for healthcare systems. In response to this need and the deficiencies of existing tools, IMVARIA Inc. will adapt its fully automated, AI-powered digital biopsy (DigitalBx) tool for idiopathic pulmonary fibrosis (IPF)—the first diagnostic AI tool for any lung disease authorized with both FDA Breakthrough Designation and novel AMA CPT billing codes, and the only such tool in active clinical use—to the non-invasive analysis of nodules suspi- cious of lung cancer from CT scans. We will provide clinicians with an end-to-end automated diagnostic solution implemented as a secure cloud-based application fully reimbursable with CPT codes, alleviating the cost and time burdens of other AI tools. We will also leverage our team's multidisciplinary expertise in AI, medical imaging, and clinical pulmonology to achieve significant, clinically impactful improvements over existing non-invasive al- ternatives. In Specific Aim 1 of this proposal, we will build, train, and test a new two-stage AI model to both identify and classify pulmonary nodules in CT scans according to risk of malignancy (ROM), training the model with ground-truth labels derived not from radiologist-labeled images (as is typical) but from broader clinical data such as surgical pathology and two-year clinical follow-up outcomes. We will optimize the model using advanced techniques such as Vision Transformers and 3D bounding spheres, with a target area under the receiver-oper- ating characteristic curve (AUROC) of at least 0.95, which would be a substantial improvement over standard CT or PET/CT assessment, while maintaining or improving sensitivity. In Specific Aim 2, we will bring the product closer to FDA approval by conducting a validation study with the optimized and locked system in accordance with CADe/CADx predicate pathway requirements. We will engage expert radiologists and pulmonologists from Mayo Clinic in a reader study to assess ROM with and without the AI tool, with the goal of demonstrating a statistically significant absolute improvement in AUROC of at least 5% with the AI tool. Upon achieving these milestones, IMVARIA plans to seek Phase II SBIR funding for a larger-scale prospective validation study with Mayo Clinic, ultimately obtaining FDA approval for DigitalBx in the diagnosis of suspicious pulmonary nodules.