Beyond Chatbot: Transforming Cancer Diagnosis and Treatment Through GenAI and Precision Transformers
Please join us on Thursday, April 30, 2026, at 12:00pm-1:00pm for the Biostatistics & Bioinformatics Core's Webinar. Dongxiao Zhu, PhD, Professor of the Department of Computer Science at Wayne State University, Director at the Institute for AI and Data Science, and Director at the Trustworthy AI Lab, will be giving a presentation on, "Beyond the Chatbot: Transforming Cancer Diagnosis and Treatment Through Generative AI and Precision Transformers" This is a virtual only seminar. Registration is required. You may register for free using the link below.
Register for free using the following link:
*Upon registration, you will receive a confirmation email with information about joining the meeting.
Abstract: Clinicians are increasingly challenged by the scale and heterogeneity of clinical data—from sparsely labeled imaging to fragmented patient records. This talk presents a unified framework that integrates Generative AI and Vision Transformers to bridge the gap from data understanding to high-precision clinical action. We begin by examining Large Language Models (LLMs) as clinical assistants for distilling complex patient histories, supporting automated consultations, and enabling personalized, interpretable risk assessment. To address critical concerns around HIPAA compliance and black-box hallucinations, we introduce white-box AI approaches that provide secure, localized, and interpretable predictions through attention-based feature attribution.
The second part focuses on precision imaging, where transformer-based architectures and foundation models address the label scarcity bottleneck in cancer segmentation. We highlight three key innovations: (1) FocalUNETR, which captures local–global interactions for boundary-aware segmentation in low-contrast CT; (2) AutoProSAM, which enables automated 3D multi-organ segmentation without manual prompting; (3) MulModSeg, which improves unpaired multi-modal (CT/MR) segmentation via modality-conditioned text embeddings. Finally, we present FluenceFormer, a physics-informed framework for automated radiotherapy planning that predicts clinically deliverable fluence maps. This end-to-end “assessment-to-action” pipeline illustrates how AI can evolve from a diagnostic support tool to a clinically actionable system for precise and reliable cancer treatment.