RETURN TO ABSTRACT LISTING


Background
AI and machine-learning applicability in radiologic imaging have historically been narrow in scope (i.e., limited to detection of a single disease state in a single organ) or restricted to the evaluation of binary states (e.g., malignant vs. non-malignant, high vs. low grade). Furthermore, robust development of deep learning requires sufficient annotated data, which can require cases numbering in the tens of thousands or more. The conventional approach of manual annotation by radiologists is impractical at that scale and is currently lacking. Lastly, a lack of rigorous demonstrations of external validity on independent datasets often hinders widespread adoption of so-called “in-house” decision classifiers. The present study aims to address these limitations by applying a bidirectional training algorithm between 2 major health systems using federated learning to increase the generalizability of a previously developed multi-organ multi-disease CNN detection system.

Educational Goals / Teaching Points
AI and machine learning can be implemented to screen for high-risk lesions in the chest, abdomen, and pelvis. Weakly supervised machine learning can help bolster sample sizes and allow for the construction of more robust classifier algorithms. Federated learning has become a necessary solution to the concerns around protected health information since only model parameters can be exchanged with private data never leaving a site. External validation is an essential precursor toward prospective implementations into clinical workflows.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
Our study aims to conduct an external validation of a previously developed “triage system” capable of broadly characterizing lesions of the chest, abdomen, and pelvis on CT scans using 3D CNNs. A prior study by a laboratory at our partner institution broadly characterized abnormalities on 13,667 body CT scans within a major health system using a proprietary lesion detection algorithm. The present study aims to demonstrate the external validity of the algorithm developed by our collaborators by expanding on the original dataset to include a massive, independent testing cohort comprised of clinical data collected by another major health system. Federated learning strategies will aid in the development of global and personalized models to train the triage system without protected data leaving a given site. Separate models are first trained separately on each institution. Each model’s performance is evaluated on the local internal and external validation from that institution. Local CNN models will be tested using our collaborator’s cohort to demonstrate bidirectional compatibility and assess the generalizability of performance.

Conclusion
This analysis will reveal the generalizability of machine learning algorithms developed and trained at a single center and subsequently applied to imaging data acquired at other institutions. Such demonstrations of generalizability are necessary steps to moving toward prospective implementations of AI algorithmic detection systems and will hopefully aid in further solidifying a role for such systems in future clinical practice.