Periodic Reporting for period 1 - BosomShield (A comprehensive CAD system based on radiologic- and pathologic-image biomarkers for diagnosis and prognosis of breast cancer relapse)
Reporting period: 2022-09-01 to 2024-08-31
Breast cancer, as one of the most prevalent cancers affecting women globally, poses significant challenges to public health systems worldwide. Early detection breast cancer relapse and precise prognostic assessment are critical for enhancing patient survival rates and quality of life. Funded by the European Union’s Horizon Europe Programme, the BosomShield project is at the forefront of combating breast cancer by harnessing innovative computational technologies and fostering international collaboration.
Main Objective: The core objective of the BosomShield project is to create a sophisticated, cloud-based Computer-Aided Diagnosis (CAD) system that uniquely integrates both histopathological and radiological imaging analyses through the use of advanced artificial intelligence (AI) techniques. This innovative integration aims to enhance the precision of breast cancer detection and the assessment of prognosis, providing clinicians with robust tools to predict breast cancer relapse and personalize treatment strategies for better patient outcomes.
Specific Objectives:
1) BosomShield is a pioneer in combining histopathological images obtained from Whole Slide Imaging (WSI) with traditional radiological imaging techniques such as mammography, MRI, and ultrasound. This multimodal approach allows for a more comprehensive analysis of cancer tissues, offering a deeper insight into the disease's characteristics.
2) In response to growing data privacy concerns and stringent regulations, BosomShield employs federated learning. This approach allows the AI models to learn from diverse, decentralized data sources without actual data transfer, ensuring patient confidentiality and data integrity.
3) The CAD system's development is guided by continuous feedback from oncologists and radiologists to ensure its alignment with real-world clinical workflows. The focus is on creating an intuitive user interface that is easy to adopt and facilitates efficient usage in clinical settings.
4) The project not only focuses on technological advancement but also on nurturing talent. Doctoral Candidates involved in BosomShield receive interdisciplinary training, equipping them with the necessary skills to spearhead future breakthroughs in cancer diagnosis, prognosis and treatment.
Impact and Vision: BosomShield aims to transform the management of breast cancer care by significantly improving diagnostic accuracy and prognostic evaluations. This project is set to transform treatment decisions, leading to enhanced therapeutic outcomes and improved quality of life for patients. Through its innovative integration of WSI with radiological imaging and its collaborative approach, BosomShield is poised to set new standards in the fight against breast cancer.
Main Objective: The core objective of the BosomShield project is to create a sophisticated, cloud-based Computer-Aided Diagnosis (CAD) system that uniquely integrates both histopathological and radiological imaging analyses through the use of advanced artificial intelligence (AI) techniques. This innovative integration aims to enhance the precision of breast cancer detection and the assessment of prognosis, providing clinicians with robust tools to predict breast cancer relapse and personalize treatment strategies for better patient outcomes.
Specific Objectives:
1) BosomShield is a pioneer in combining histopathological images obtained from Whole Slide Imaging (WSI) with traditional radiological imaging techniques such as mammography, MRI, and ultrasound. This multimodal approach allows for a more comprehensive analysis of cancer tissues, offering a deeper insight into the disease's characteristics.
2) In response to growing data privacy concerns and stringent regulations, BosomShield employs federated learning. This approach allows the AI models to learn from diverse, decentralized data sources without actual data transfer, ensuring patient confidentiality and data integrity.
3) The CAD system's development is guided by continuous feedback from oncologists and radiologists to ensure its alignment with real-world clinical workflows. The focus is on creating an intuitive user interface that is easy to adopt and facilitates efficient usage in clinical settings.
4) The project not only focuses on technological advancement but also on nurturing talent. Doctoral Candidates involved in BosomShield receive interdisciplinary training, equipping them with the necessary skills to spearhead future breakthroughs in cancer diagnosis, prognosis and treatment.
Impact and Vision: BosomShield aims to transform the management of breast cancer care by significantly improving diagnostic accuracy and prognostic evaluations. This project is set to transform treatment decisions, leading to enhanced therapeutic outcomes and improved quality of life for patients. Through its innovative integration of WSI with radiological imaging and its collaborative approach, BosomShield is poised to set new standards in the fight against breast cancer.
Technical and Scientific Activities:
Robust methods were developed for tumor detection, segmentation, and molecular subtype classification using radiological images (mammograms, MRI, PET). High accuracy was achieved in tumor localization (~98%) and molecular subtype correlation (~88%) using deep learning. Innovative segmentation models (cGANs) enhanced breast density analysis, achieving 94% accuracy and aiding early BC diagnosis and recurrence prediction.
Whole-Slide Imaging (WSI) processing was standardized. Biomarkers were developed for primary and axillary tumors, using immune biomarker localization and morphology to predict metastases risk. Explainable AI (XAI) identified critical histopathological features, improving interpretability.
Clinical data was integrated with radiological and histopathological findings to enhance BC relapse prediction models. Imaging biomarkers were correlated with clinical data for personalized risk assessment.
A privacy-preserving Federated Learning (FL) framework was established to train AI models across decentralized datasets, complying with GDPR. Threat models and countermeasures (e.g. MemberShield) were developed to safeguard data integrity and prevent attacks.
A Collaborative and Secure CAD System for BC relapse prediction was designed and implemented, leveraging cloud computing with built-in security and privacy. Preliminary versions showed promise in harmonizing multi-modal data.
XAI models were integrated into pipelines for interpretable predictions. Feature importance analyses validated model behavior.
Extensive model validation was conducted using public and private datasets to ensure generalizability. Results highlighted model robustness in diverse imaging and clinical scenarios.
Key Outcomes:
1. High-performing, interpretable AI models for BC detection and relapse prediction.
2. Federated Learning protocols for secure, compliant multi-institutional collaboration.
3. A CAD system framework integrating multi-modal data for advanced decision support.
4. Pioneering methodologies for bias mitigation, ethical AI, and reproducibility.
These achievements underscore BosomShield's advancements, setting a strong foundation for future progress.
Robust methods were developed for tumor detection, segmentation, and molecular subtype classification using radiological images (mammograms, MRI, PET). High accuracy was achieved in tumor localization (~98%) and molecular subtype correlation (~88%) using deep learning. Innovative segmentation models (cGANs) enhanced breast density analysis, achieving 94% accuracy and aiding early BC diagnosis and recurrence prediction.
Whole-Slide Imaging (WSI) processing was standardized. Biomarkers were developed for primary and axillary tumors, using immune biomarker localization and morphology to predict metastases risk. Explainable AI (XAI) identified critical histopathological features, improving interpretability.
Clinical data was integrated with radiological and histopathological findings to enhance BC relapse prediction models. Imaging biomarkers were correlated with clinical data for personalized risk assessment.
A privacy-preserving Federated Learning (FL) framework was established to train AI models across decentralized datasets, complying with GDPR. Threat models and countermeasures (e.g. MemberShield) were developed to safeguard data integrity and prevent attacks.
A Collaborative and Secure CAD System for BC relapse prediction was designed and implemented, leveraging cloud computing with built-in security and privacy. Preliminary versions showed promise in harmonizing multi-modal data.
XAI models were integrated into pipelines for interpretable predictions. Feature importance analyses validated model behavior.
Extensive model validation was conducted using public and private datasets to ensure generalizability. Results highlighted model robustness in diverse imaging and clinical scenarios.
Key Outcomes:
1. High-performing, interpretable AI models for BC detection and relapse prediction.
2. Federated Learning protocols for secure, compliant multi-institutional collaboration.
3. A CAD system framework integrating multi-modal data for advanced decision support.
4. Pioneering methodologies for bias mitigation, ethical AI, and reproducibility.
These achievements underscore BosomShield's advancements, setting a strong foundation for future progress.
BosomShield developed robust methods for tumor detection, segmentation, and molecular subtype classification from radiological images (mammograms, MRI, PET), achieving high accuracy (~98% localization, ~88% subtype correlation) using deep learning and innovative cGANs for breast density analysis (94% accuracy).
Standardized WSI processing enabled biomarker development for primary/axillary tumors to predict metastases risk using immune biomarker localization and morphology, with XAI improving interpretability.
Clinical data integration enhanced BC relapse prediction models and personalized risk assessment.
A privacy-preserving Federated Learning framework (GDPR compliant) with threat models (e.g. MemberShield) was established. A secure, cloud-based CAD system for BC relapse prediction, harmonizing multi-modal data, showed promise.
XAI models were integrated for interpretable predictions, validated by feature importance analyses.
Extensive validation across diverse datasets demonstrated model robustness.
Key outcomes include high-performing AI models, Federated Learning protocols, a multi-modal CAD system framework, and pioneering methodologies for bias mitigation, ethical AI, and reproducibility, setting a strong foundation for future progress.
Standardized WSI processing enabled biomarker development for primary/axillary tumors to predict metastases risk using immune biomarker localization and morphology, with XAI improving interpretability.
Clinical data integration enhanced BC relapse prediction models and personalized risk assessment.
A privacy-preserving Federated Learning framework (GDPR compliant) with threat models (e.g. MemberShield) was established. A secure, cloud-based CAD system for BC relapse prediction, harmonizing multi-modal data, showed promise.
XAI models were integrated for interpretable predictions, validated by feature importance analyses.
Extensive validation across diverse datasets demonstrated model robustness.
Key outcomes include high-performing AI models, Federated Learning protocols, a multi-modal CAD system framework, and pioneering methodologies for bias mitigation, ethical AI, and reproducibility, setting a strong foundation for future progress.