Design and development of an AI algorithm to predict recurrence and treatment response in NSCLC patients from CT and pathology images
Principal Investigator
Name
Raffaella Massafra
Degrees
M.Sc
Institution
Istituto Tumori "Giovanni Paolo II" I.R.C.C.S.
Position Title
Head of the Departmental Unit of Medical Physics at Istituto Tumori "Giovanni Paolo II" I.R.C.C.S.
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1269
Initial CDAS Request Approval
Jun 17, 2024
Title
Design and development of an AI algorithm to predict recurrence and treatment response in NSCLC patients from CT and pathology images
Summary
The prognosis of non-small cell lung cancer (NSCLC) is poor, with a 5-year survival rate ranging from 68% for patients in stage IB to 0-10% for patients in stages IVA-IVB. Early assessment of the likelihood of a patient's response to a specific treatment is necessary to evaluate the utility or futility of the treatment in the medium and long term, considering potential toxicities.
In this study we hypothesize that a radiomic signature obtained from CT images of patients with lung pathology can predict the response to therapy and the treatment-related toxicity. The gold standard for treating unresectable stage III NSCLC is represented by chemoradiotherapy (CRT), and the addition of immunotherapy (IT) has further improved patient outcomes. One of the major limitations in clinical practice is the timing of radiological re-evaluation after CRT, a fundamental requirement to start IT in PD-L1 positive patients who have not progressed.
The morphological characteristics of the disease extracted from CT scans and digital slides of tumor biopsies, known as digital pathology images, using artificial intelligence (AI) algorithms, have the potential to be used as predictive markers of response in patients with unresectable NSCLC undergoing definitive CRT. Hence, the idea of using this approach to predict the response to CRT treatment.
The research project has multiple objectives, including statistical evaluations of associations between radiomic features extracted from pre-treatment CT images and digital pathology with clinical characteristics and patient outcomes.
Integration of information from multiple sources (e.g., clinical and imaging data) and their practical use as decision support will be carried out through multivariate statistical models, AI algorithms, including machine learning and deep learning techniques, particularly standard classifiers (Random Forest, Support Vector Machine) or Artificial Neural Networks. Feature selection techniques will be nested within classification models to identify the most significant features for prediction activities. Standard metrics, such as accuracy, sensitivity, and specificity, as well as the Area Under the Curve (AUC), will be calculated to evaluate model performance using cross-validation schemes and independent tests.
For each individual patient, the implemented AI model will make a choice in accordance with the predicted outcome. This choice will be clarified by explainable AI methods, such as SHapley Additive exPlanations (SHAP), based on the calculation of Shapley values, and Local Interpretable Model-agnostic Explanations (LIME). Each patient will be represented by an importance vector, where each element will represent the weight of how each feature, considered alone or in collaboration with all other considered features, will contribute to the AI model's final choice.
In this study we hypothesize that a radiomic signature obtained from CT images of patients with lung pathology can predict the response to therapy and the treatment-related toxicity. The gold standard for treating unresectable stage III NSCLC is represented by chemoradiotherapy (CRT), and the addition of immunotherapy (IT) has further improved patient outcomes. One of the major limitations in clinical practice is the timing of radiological re-evaluation after CRT, a fundamental requirement to start IT in PD-L1 positive patients who have not progressed.
The morphological characteristics of the disease extracted from CT scans and digital slides of tumor biopsies, known as digital pathology images, using artificial intelligence (AI) algorithms, have the potential to be used as predictive markers of response in patients with unresectable NSCLC undergoing definitive CRT. Hence, the idea of using this approach to predict the response to CRT treatment.
The research project has multiple objectives, including statistical evaluations of associations between radiomic features extracted from pre-treatment CT images and digital pathology with clinical characteristics and patient outcomes.
Integration of information from multiple sources (e.g., clinical and imaging data) and their practical use as decision support will be carried out through multivariate statistical models, AI algorithms, including machine learning and deep learning techniques, particularly standard classifiers (Random Forest, Support Vector Machine) or Artificial Neural Networks. Feature selection techniques will be nested within classification models to identify the most significant features for prediction activities. Standard metrics, such as accuracy, sensitivity, and specificity, as well as the Area Under the Curve (AUC), will be calculated to evaluate model performance using cross-validation schemes and independent tests.
For each individual patient, the implemented AI model will make a choice in accordance with the predicted outcome. This choice will be clarified by explainable AI methods, such as SHapley Additive exPlanations (SHAP), based on the calculation of Shapley values, and Local Interpretable Model-agnostic Explanations (LIME). Each patient will be represented by an importance vector, where each element will represent the weight of how each feature, considered alone or in collaboration with all other considered features, will contribute to the AI model's final choice.
Aims
- Develop a meaningful radiomic signature to predict the response to therapy and the treatment-related toxicity in patients with stage III and IV NSCLC;
- Conduct statistical analyses to explore the relationships between radiomic features extracted from pre-treatment CT images and digital pathology with clinical characteristics and patient outcomes;
- Develop an automated and personalized system for planning and monitoring therapeutic efficacy based on radiomic analysis of radiological and histopathological images of the neoplasm.
Collaborators
Annarita Fanizzi, Istituto Tumori "Giovanni Paolo II" I.R.C.C.S.
Maria Colomba Comes, Istituto Tumori "Giovanni Paolo II" I.R.C.C.S.
Samantha Bove, Istituto Tumori "Giovanni Paolo II" I.R.C.C.S.
Arianna Campione, Istituto Tumori "Giovanni Paolo II" I.R.C.C.S.