EE ZOOM Seminar: Automated Computerized Analysis of Pulmonary Embolism Prognosis using Multimodal Deep Learning Diagnostic Tools

https://tau-ac-il.zoom.us/j/82455834599?pwd=4MbO0LWdYoRespFabd9sGSTuGUNi6y.1

Meeting ID: 824 5583 4599

 Passcode: 291308

Electrical Engineering Systems ZOOM Seminar

Speaker: Noa Cahan

Ph.D. student under the supervision of Prof. Hayit Greenspan

Sunday, 29^th June 2025, at 15:00

Automated Computerized Analysis of Pulmonary Embolism Prognosis using Multimodal Deep Learning Diagnostic Tools

Abstract

In this research, we focus on deep learning applications for the detection, delineation, and risk stratification of pulmonary embolism (PE) - a critical, life-threatening condition. Rapid and accurate risk stratification can decrease PE mortality rates. Computed Tomography Pulmonary Angiography (CTPA) is the gold standard diagnostic tool. Unlike previous research that predominantly relies on imaging-only approaches for PE clot detection, our work innovatively integrates various data modalities, enhancing the accuracy and efficiency of risk assessments. The data includes: (1) Imaging Data: Computed Tomography Pulmonary Angiography (CTPA), Chest X-Rays, and Electrocardiograms. (2) Tabular Clinical Data: demographics, comorbidities, vital signs, laboratory results, and clinical scores. The research addresses prevalent challenges including limited annotated data, biases in existing models, and ensuring robustness in the developed AI tools. A significant emphasis is placed on explainability in AI models, ensuring transparency and trust in medical decision-making processes. Further, motivated by recent advancements in generative AI, we intend to pioneer the use of few-shot learning and cross model transfer, applying diffusion models to transform 2D-X-rays into 3D-CTPA equivalents. The generated 3D-CTPA, can later be used for PE classification. Success in this endeavor could potentially eliminate the need for CTPA scans.

To our knowledge, no prior studies have automated PE severity assessment or used diffusion models for X-ray to CT conversion. Our solutions aim to expedite diagnosis, enhance treatment times, and refine risk assessments. The tasks defined could improve the ability to direct preventative and health surveillance resources and advance healthcare as a whole. The specific papers we will present in this talk include:

1. Multimodal fusion models for pulmonary embolism mortality prediction.
2. X-ray2CTPA: Leveraging Diffusion Models to Enhance Pulmonary Embolism Classification.
3. Cross-Modal CXR-CTPA Knowledge Distillation using latent diffusion priors towards CXR Pulmonary Embolism Diagnosis.