Predicting tumor evolution during radiotherapy is a clinically critical challenge, particularly when longitudinal changes are driven by both anatomy and treatment. In this work, we introduce a Virtual Treatment (VT) framework that formulates non-small cell lung cancer (NSCLC) progression as a dose-aware multimodal conditional image-to-image translation problem. Given a CT scan, baseline clinical variables, and a specified radiation dose increment, VT aims to synthesize plausible follow-up CT images reflecting treatment-induced anatomical changes. We evaluate the proposed framework on a longitudinal dataset of 222 stage III NSCLC patients, comprising 895 CT scans acquired during radiotherapy under irregular clinical schedules. The generative process is conditioned on delivered dose increments together with demographic and tumor-related clinical variables. Representative GAN-based and diffusion-based models are benchmarked across 2D and 2.5D configurations. Quantitative and qualitative results indicate that diffusion-based models benefit more consistently from multimodal, dose-aware conditioning and produce more stable and anatomically plausible tumor evolution trajectories than GAN-based baselines, supporting the potential of VT as a tool for in-silico treatment monitoring and adaptive radiotherapy research in NSCLC.
Alex: Welcome to another episode of ResearchPod.
Sam: Today, we're looking at a study from researchers at Università Campus Bio-Medico di Roma and others. It explores a central puzzle: how artificial intelligence can predict changes in non-small cell lung cancer—NSCLC, the most common type—during radiotherapy, when follow-up scans happen at irregular times.
Alex: So tumors change between scans, but follow-up CT scans don't happen on a regular schedule?
Sam: Yes. Patients start with a planning CT scan. Follow-ups come weeks apart due to side effects or hospital schedules. This leaves doctors unable to see tumor shrinkage or shifts until after more radiation doses. It's a blind spot in treatment planning.
Alex: The radiation dose itself drives those changes—like targeted energy damaging cancer cells. Without timely scans, adjustments lag.
Sam: Exactly. NSCLC tumors move with breathing, and lung tissue shifts in complex ways. Past AI methods ignored the radiation dose when predicting scans. This study introduces Virtual Treatment, or VT. It generates a predicted future CT scan from the current one, plus patient details like age and tumor stage, and the extra dose since the last scan. Think of it like fast-forwarding a video: you input the starting frame and how much "push" to add, and it shows the next frame.
Alex: A dose increment is just the added radiation. Why focus on that?
Sam: It lets the AI learn real treatment effects without relying on irregular scan timing. They trained on data from 222 stage III NSCLC patients.
Alex: Can AI really predict tumor changes based on dose alone, filling those gaps?
Sam: The paper suggests yes. They call it dose-aware image translation. Picture turning one photo into a future version, guided by the current scan, patient facts, and dose. It blends image and numbers to simulate tumor response on a computer—no new physical scans needed.
Alex: Doctors could ask, "What happens with 10 more gray?" and get a virtual CT instantly. That speeds up tweaking plans based on response.
Sam: Right. They trained by pairing each patient's earlier and later scans, matching them to the dose in between. This handles any dose at prediction time.
Alex: Patients differ by age, sex, and staging—how far cancer has spread. How do predictions stay realistic?
Sam: They tested AI generators and found diffusion models best—like their adapted version. These start with random noise, like TV static, and gradually sharpen it into a clear image, guided by inputs. It's steadier for tracking chest changes over time than older methods.
Alex: They also focus training on the tumor area?
Sam: Yes. They target the Clinical Target Volume—the tumor plus a safety margin from the starting scan. Training penalizes errors only there, not the whole image. This prioritizes the key spot where radiation hits.
Alex: No need to re-outline the tumor each time. Do predictions better capture shrinkage?
Sam: The evidence points that way. Volume checks show errors under 25%—clinically acceptable—even at higher doses. Visuals confirm diffusion shows steady tumor regression without glitches.
Alex: Diffusion takes more computing power, but the study notes costs drop after training. VT could let doctors test dose plans on a computer first.
Sam: Yes. It closes the gap between scans, aiding adaptive radiotherapy by simulating responses quickly.
Alex: A practical step forward in tracking treatment effects. Thanks for joining us on ResearchPod.