Selected works

Reliable AI in
radiation oncology.

Selected research by Weizhi Du and collaborators in radiation oncology, alongside projects in multi-agent planning, reinforcement learning, and computer vision.

Research

Manuscripts, chapters & conference abstracts.

Safety-aware segmentation workflow, validation guardrails, and results

01Sole first author

A Safety-Aware Agentic AI System for Refining FDG-PET Auto-Contouring in Locally Advanced Cervical Cancer

72nd Annual Meeting of the Radiation Research Society (RRS), 2026.

Bounded, auditable refinement of a MONAI segmentation baseline, with validation gates and clinician oversight.

Read abstract ↗

Diffusion denoising, transformer attention, and autoregressive language generation

02Sole first author

Generative AI and Large Language Models

In Generative AI Models and Reinforcement Learning in Radiotherapy Treatment Planning, Chapter 3. CRC Press, forthcoming.

Diffusion models, transformers, language models, and their clinical evaluation in radiotherapy planning.

Read chapter ↗

Benchmark accuracy for four model generations across no-prompt, baseline, RAG-only, and eRT-RAG conditions

03Co-author

Evaluation of Self-Evolving Multi-Agent Retrieval-Augmented Generation in Radiation Oncology

JAMA Oncology, 2026. Manuscript under review.

Evidence-grounded reasoning across four model generations, board-style questions, and expert-rated clinical debates.

Read manuscript ↗

eRT-RAG retrieval architecture and the feedback-driven knowledge update loop

04Co-author

Self-Evolving Agentic Retrieval-Augmented Generation Enables High-Accuracy and Confidence-Aware Radiation Oncology Reasoning

72nd Annual Meeting of the Radiation Research Society (RRS), 2026.

Agentic retrieval, confidence-aware validation, and a continuously updated radiation oncology knowledge base.

Read abstract ↗

Projects

Methods, experiments & technical reports.

Agent contacts per task across six acquisition methods, with bars for means and dots for validated trials

P01

Budgeted Continual Multi-Agent Planning with Hidden Capabilities

Selective capability discovery reduced agent contacts by 61% versus random querying and action-schema disclosure by 62% versus Reveal-All, across 576 validated trials.

Read project report ↗

Evaluation win-rate curves for the baseline and pruned Gomoku models against two opponents

P02

Making Self-Play Reinforcement Learning Feasible for Gomoku via Candidate Move Restriction

Threat-aware MCTS pruning completed 8×8 connect-5 training in 7 hours, versus an unfinished 48-hour unrestricted run, and reached a 60% peak win rate against pure MCTS.

Read project report ↗

The query “cat right” localizes the kitten on the right, with ranked boxes and a ground-truth annotation

P03

Seeing What You Mean: Language-Guided Scene Understanding with Explanation

A training-free visual grounding pipeline improved Acc@0.5 from 40.6% to 62.8%, with 0.614 mIoU across 5,598 validation queries.

Read project report ↗

For questions or an updated copy, contact Weizhi Du ↗.

Reliable AI inradiation oncology.

Research

A Safety-Aware Agentic AI System for Refining FDG-PET Auto-Contouring in Locally Advanced Cervical Cancer

Generative AI and Large Language Models

Evaluation of Self-Evolving Multi-Agent Retrieval-Augmented Generation in Radiation Oncology

Self-Evolving Agentic Retrieval-Augmented Generation Enables High-Accuracy and Confidence-Aware Radiation Oncology Reasoning

Projects

Budgeted Continual Multi-Agent Planning with Hidden Capabilities

Making Self-Play Reinforcement Learning Feasible for Gomoku via Candidate Move Restriction

Seeing What You Mean: Language-Guided Scene Understanding with Explanation

Reliable AI in
radiation oncology.