AI-based multi-dimensional characterization of pulmonary nodules to improve lung cancer screening
However, despite improved consistency, persistent inter-grader variability and incomplete image characterization remain as limitations. Furthermore, the nodule’s profiling is often based on a single time point, disregarding any additional information that could come from the disease progression. Finally, adding information coming from different modalities, for example histopathology, could provide a more detailed and comprehensive characterization of the disease. Enhancing the sensitivity and specificity of lung cancer screening is imperative to improve the clinical outcome and lower the financial costs.
Automated methods, e.g. employing Artificial Intelligence (AI) algorithms, have the potential to overcome those limitations, leading to a reduction in the burden on radiologists and a decrease in variability
This project aims thus to enhance the cost-effectiveness and facilitate personalized care in lung cancer screening programs using LDCT and pathology images. Using both cross-sectional and longitudinal data, we aim to develop a new AI-based method able to perform an accurate risk-stratification of patients affected by LC and to provide a more personalized prediction of patients’ mortality.
1) Apply AI methods to detect patients with at least one nodule within the cross-sectional LDCT image of the lung.
2) Segment CT images from positive patients to isolate suspicious nodules.
3) Extract quantitative and deep features from each nodule to perform risk-stratification of patients affected by LC. The effect of additional comorbidity factors, such as chronic obstructive pulmonary disease and cardiovascular disease, will be also investigated.
4) Develop and evaluate different image registration approaches (both intensity and feature-based) to reliably align lung LDCT acquired longitudinally over time.
5) Implement image subtraction pipelines to identify those nodules that evolve over time and new nodules that appear in follow up time points.
6) Perform an early classification of evolving nodules from baseline LDCT.
7) Include information about the longitudinal nodule evolution and new nodules in the model in 3) to investigate whether it could provide additional value to risk stratification compared to cross-sectional approach.
8) Use AI to segment cell’s nuclei from histopathology images to identify abnormal cell features, such as nuclear atypia and abundant tumor cellularity.
9) Include histopathology-related features in the predictive model to evaluate their impact in the risk stratification.
Dr. Marco Battaglini, Siena Imaging SRL
Dr. Barbara Iadarola, Siena Imaging SRL
Dr. Ludovico Luchetti, Siena Imaging SRL
Dr. Giordano Gentile, Siena Imaging SRL
Dr. Giacomo Demurtas, Siena Imaging SRL
Prof. Stefano Diciotti, Department of Electrical, Electronic, and Information Engineering "Guglielmo Marconi" (DEI), Alma Mater Studiorum - University of Bologna
Dr. Giulia Raffaella De Luca, Department of Electrical, Electronic, and Information Engineering "Guglielmo Marconi" (DEI), Alma Mater Studiorum - University of Bologna