Building integrative multi-modal and interpretable AI systems for personalized lung cancer therapy
This research aims to develop Artificial Intelligence (AI) based systems by integrating multi-omic data for personalized therapy in Non-Small Cell Lung Cancer (NSCLC). We propose a novel deep learning-enabled framework to extract image biomarkers tumor microenvironment on H&E-stained histopathology images. Also, we will establish robust radiomic features on Computerized Tomography (CT) images. Finally, we will integrate radiomic, pathomic, clinical, and molecular phenotyping data to provide a holistic and multi-modal analytic approach to predict lung cancer recurrence.
Clinical background:
Lung cancer is the leading cause of cancer-related deaths in the United States and worldwide. NSCLC accounts for the vast majority (~85%) of lung cancers. The current standard of care for nonmetastatic stage I-III NSCLC is curative-intent treatment with surgery and/or radiotherapy with or without adjuvant systemic therapy. Currently, the decision of whether to treat early-stage NSCLC with adjuvant therapy is based on clinicopathologic criteria, i.e., tumor size ≥ 4cm or lymph node involvement. However, current prognostic factors are rather crude. There is a critical unmet need for biomarkers that reliably predict which cancers will recur for an individual patient.
Aim 1: Develop pathological biomarker leveraging deep learning approaches
We will develop a computational pipeline to extract relevant deep image features from tumor microenvironment on H&E stained images. We will then use machine learning to visualize distinctive histopathologic features detected by the machine learning model and confirm via molecular phenotyping.
Aim 2: Here, we will prospectively evaluate and establish robust radiomic features of lung cancer based on CT images.
Aim 3: Develop integrative multi-modal biomarkers
Finally, we will build integrative models by combining different data modalities including clinical, pathomic, radiomic, and molecular phenotyping data to predict lung cancer recurrence. Here, we will investigate feature fusion as well as decision fusion.
Rasoul Sali, Postdoctoral scholar in Radiation Oncology, Stanford University