Improving lung cancer risk prediction with artificial intelligence
The NLST has demonstrated that screening high-risk patients with LDCT leads to a mortality reduction from lung cancer of more than 20%.
Despite the potential benefit of LDCT, the false positive rate is high due to the high prevalence of indeterminate pulmonary nodules and the limited quality of LDCT. The subsequent overdiagnosis results in further unnecessary, invasive treatments to patients, leading to an excessive overload for the health system. These aspects, along with the radiologists’ reading procedures which are time consuming and operator dependent, limit the implementation of prevention programs.
Artificial Intelligence (AI) based strategies, such as those based on radiomics and machine learning (ML) demonstrated to be promising in determining nodule’s probability of malignancy, however its reliability still needs to be increased to change the current standards of care. Our intent is to develop a Computer Assisted Diagnosis system based on AI able to handle pulmonary nodules and support decision making in lung cancer screening programs, limiting invasive unnecessary examinations, and reducing the amount of time spent by radiologists on each medical exam. We are willing to improve the quality of LDCT imaging data and implement AI-based strategies for lesion detection and future appearance prediction in terms of malignancy and growth rate.
Moreover, we aim to investigate the internal logic of AI models through Explainable AI (XAI) tools, since their ‘black-box’ approach is still a major concern preventing their introduction in the clinical practice.
To achieve our objectives, we are going to focus on three main activities: (i) the improvement of the quality of LDCT imaging data, (ii) the implementation of predictive AI models to estimate lung cancer risk from LDCT images, and (iii) the technical evaluation of the explainability of the models, supported by a clinical and biological validation.
• Firstly, we want to implement deep learning-based image reconstruction (DLIR) methods to improve the quality of LDCT images and perform a comparison with different analytical reconstruction algorithms.
• Secondly, we plan to utilize radiomics to extract quantitative image biomarkers to be integrated in ML models able to predict the malignancy of the identified lesions. Deep learning methods will also be investigated to skip features extraction. Models able to predict tumor growth are also foreseen to be implemented, to provide indications on time for follow-up exams and, eventually, surgery.
• Finally, to address the current lack of interpretability of the AI models, a further analysis will be carried out on the results to explain the causal relationship between input features and outcome from a technical point of view, along with the support of clinical and biological information available in the NLST database.
The LDCT images available in the NTLS dataset would allow us to develop the presented work path. It perfectly suits our intentions because the clinical information collected in the dataset, such as clinical scores of tumor malignancy and pathological exams, would allow us to perform, respectively, a clinical and biological validation of the models.
Chiara Paganelli:
Assistant Professor at Politecnico di Milano
chiara.paganelli@polimi.it
Francesca Camagni:
Ph.D. student in Bioengineering at Politecnico di Milano
francesca.camagni@polimi.it
Mariagrazia Monteleone:
Ph.D. student in Bioengineering at Politecnico di Milano
mariagrazia.monteleone@polimi.it