Robust and reliable Artificial Intelligence for oncology
Principal Investigator
Name
Chiara Verdone
Degrees
Ph.D.
Institution
Università degli Studi del Sannio
Position Title
Ph.D.
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-851
Initial CDAS Request Approval
Nov 3, 2021
Title
Robust and reliable Artificial Intelligence for oncology
Summary
Cancer is the deadliest disease of all, no matter what type of malignancy it is. All researchers and oncologists agree on the fact that early detection of cancer increases the patient's chances of survival tremendously. Unfortunately, most cancer patients are diagnosed in the final stages of the disease. Recently, Artificial intelligence (AI) and deep learning have shown major significance in solving this issue.
Diagnostic imaging plays an integral role in many phases of a patient’s care path, but considerable challenges remain regarding accuracy, efficacy, and efficiency. AI holds great promise to support image analysis to detect, characterize, and monitor the disease. AI can enable automated segmentation and support in diagnosis and staging. Furthermore, it can support cancer monitoring by capturing image features over time to evaluate the patient’s response to treatment.
The aim of this study is to build a hybrid model for the localization and diagnosis of cancer nodules on ultrasound and to evaluate their diagnostic performance.
The project activities are organized in 3 tightly interlinked phases:
1. Machine Learning for robust & trustworthy AI
This activity aims to develop robust and trustworthy AI in medical imaging. It will address a subset of clinical application areas (e.g., lung cancer and pancreas cancer), by developing effective and reliable AI methods for the selected clinical use cases.
Research topics include, but are not limited to, AI transparency, interpretability and explainability, self-critical AI, confidence quantification, and out-of-distribution detection.
2. Clinical data collection and heterogeneous data source analysis
The dataset will consist of both data concerning the patient's general medical history and ultrasound images.
3. Evaluation
The research comprises the evaluation of AI-based nodule assessment, evaluation of AI-based cancer detection and resectability assessment, and testing of the scalability of the innovations to a different cancer type (e.g. thyroid cancer) based on results of the previous phase.
The interest is the development of trustworthy models for medical image analysis that can deal with the enormous amount of heterogeneity in the visual representation of tumors and how we can efficiently translate AI solutions to other CT applications for oncology, e.g. from thyroid cancer to other cancers. Various types of Deep Neural Networks will be implemented and benchmarked with a special focus on explainability, uncertainty, and robustness. These aspects are key for clinical integration and an optimal collaboration between the medical specialist and the AI system, fostering the adoption of AI in clinical practice.
Diagnostic imaging plays an integral role in many phases of a patient’s care path, but considerable challenges remain regarding accuracy, efficacy, and efficiency. AI holds great promise to support image analysis to detect, characterize, and monitor the disease. AI can enable automated segmentation and support in diagnosis and staging. Furthermore, it can support cancer monitoring by capturing image features over time to evaluate the patient’s response to treatment.
The aim of this study is to build a hybrid model for the localization and diagnosis of cancer nodules on ultrasound and to evaluate their diagnostic performance.
The project activities are organized in 3 tightly interlinked phases:
1. Machine Learning for robust & trustworthy AI
This activity aims to develop robust and trustworthy AI in medical imaging. It will address a subset of clinical application areas (e.g., lung cancer and pancreas cancer), by developing effective and reliable AI methods for the selected clinical use cases.
Research topics include, but are not limited to, AI transparency, interpretability and explainability, self-critical AI, confidence quantification, and out-of-distribution detection.
2. Clinical data collection and heterogeneous data source analysis
The dataset will consist of both data concerning the patient's general medical history and ultrasound images.
3. Evaluation
The research comprises the evaluation of AI-based nodule assessment, evaluation of AI-based cancer detection and resectability assessment, and testing of the scalability of the innovations to a different cancer type (e.g. thyroid cancer) based on results of the previous phase.
The interest is the development of trustworthy models for medical image analysis that can deal with the enormous amount of heterogeneity in the visual representation of tumors and how we can efficiently translate AI solutions to other CT applications for oncology, e.g. from thyroid cancer to other cancers. Various types of Deep Neural Networks will be implemented and benchmarked with a special focus on explainability, uncertainty, and robustness. These aspects are key for clinical integration and an optimal collaboration between the medical specialist and the AI system, fostering the adoption of AI in clinical practice.
Aims
The aim of this study is to build an AI hybrid model for the localization and diagnosis of cancer nodules on ultrasound and to evaluate their diagnostic performance.
Collaborators
Lerina Aversano, University of Sannio
Mario Luca Bernardi, University of Sannio
Martina Iammarino, University of Sannio