M.O.V.I.E：應用 FiLM 調控的神經微分方程於 ¹¹C-PiB 正子斷層掃描晚期影像生成潛在動態建模研究

Abstract

我們提出 M.O.V.I.E.（Modeling ODEs with Visually Interpretable Evolution），一套可將僅約10 分鐘之早期 [¹¹C]-PiB PET 影像推估晚期（late-phase）影像。與以往將此任務視為靜態逐幀映射的方法不同，M.O.V.I.E. 將影像轉換建模為由神經常微分方程（neural ODE）或隨機微分方程（neural SDE）所支配的時間連續過程，因此能產生平滑且可解釋的示蹤劑潛在動態。為進一步貼近生理真實性，我們導入 FiLM（Feature-wise Linear Modulation）條件，利用中期 PET 訊號生理導向地調控 漂移(drift)與擴散(diffusion)過程。

與真實晚期影像相比，M.O.V.I.E. 模型在量化指標上取得顯著改善。ODE 模型將峰值信噪比 (PSNR) 自 26.69 提升至 41.23（+54.4%），SDE 模型則提高至 41.46（+55.3%），顯示重建品質大幅進步；結構相似性 (SSIM) 方面，兩者皆較基準影像提升了4.28 %。在感知品質方面，以深度特徵衡量的影像感知相似度 (LPIPS) 顯示，UNet 基線模型可將感知誤差降低 20.7 %，SDE 亦降低 17.3%。

在臨床層面，合成影像展現出強大的診斷效用。在 SDE 模型下，扣帶回（cingulate）與額葉（frontal）區域對 Aβ 陽性偵測的ROC 曲線下面積(AUC)分別達到 0.798 與 0.786；敏感性（sensitivity）提高至 63.04%，而特異性（specificity）則維持在 83.33%，使正向預測值（positive predictive value）達到 87.88%。此外，透過 PHATE 投影潛在空間，可清晰區分不同診斷群體及 Aβ 狀態的動力學模式。

M.O.V.I.E. 為首個結合時間連續性與生理引導之早期至晚期 PET 合成框架。透過跳脫傳統黑箱式迴歸（black‐box regression），轉向可解釋潛在動態，本方法展現縮短掃描時間、提升可解釋性，並於澱粉樣成像之臨床決策中提供實質助益。

We propose M.O.V.I.E. (Modeling ODEs with Visually Interpretable Evolution), a generative framework for synthesizing late-phase [¹¹C]-PiB PET images from short 1-minute (9′30′′-10′30′′ p.i.)early acquisitions. Unlike prior approaches that treat this task as static frame-to-frame mapping, M.O.V.I.E. models the transformation as a time-continuous process governed by neural ordinary or stochastic differential equations (ODEs/SDEs), enabling smooth and interpretable latent dynamics of tracer uptake. To further enhance biological realism, we introduce FiLM-based conditioning, guiding the drift and diffusion processes using mid-phase PET signals in a physiology-aware manner.

Compared to the ground-truth reference images, our M.O.V.I.E. models demonstrated substantial quantitative improvements. PSNR increased by +54.4% and +55.3% for the ODE and SDE variants, respectively, indicating superior reconstruction fidelity. In terms of structural similarity, both ODE and SDE improved SSIM by +4.28% over the reference. Furthermore, perceptual quality measured by LPIPS was significantly improved, with the UNet baseline reducing perceptual error by -20.7%, followed by -17.3% in the SDE variant. Effect size comparisons also confirmed that both ODE and SDE outperformed the UNet baseline.

At the clinical level, synthesized images yielded strong diagnostic utility. Under the SDE model, the cingulate and frontal regions achieved AUCs of 0.798 and 0.786 for amyloid-positivity detection. Sensitivity increased to 63.04%, while specificity was preserved at 83.33%, resulting in a high positive predictive value (87.88%). PHATE projections of the latent space further revealed distinct kinetic patterns across diagnostic groups and Aβ status.

M.O.V.I.E. is the first framework to bridge early-to-late PET synthesis with a temporally continuous, physiology-guided generative model. By moving beyond black-box regression toward explainable latent dynamics, it demonstrates the potential to reduce scan time, improve interpretability, and support clinical decision-making in amyloid imaging.