INet：用於磁振造影至電腦斷層影像轉換之超輕量化模型

Abstract

由磁振造影（MR）影像生成偽電腦斷層（pseudo-CT）影像，能在降低對電腦斷層（CT）掃描依賴的同時，保留治療應用所需之物理資訊。然而現有偽電腦斷層生成網路通常仰賴複雜的網路架構與高計算成本，限制其實際應用與部署。因此，本研究提出一種超輕量化且具可解釋性的偽電腦斷層生成網路INet。

相較於傳統使用的UNet，INet僅需20.5%的浮點運算量（FLOPs）以及0.58%的參數數量，即可維持相當的影像重建效能。在獨立測試集上，INet 在多項定量評估指標中皆達到與UNet及UNet3+具統計相當性的表現，並顯著優於UNeXt。定性分析結果進一步顯示，INet 能更有效保留顱骨結構連續性、皮質骨邊界以及局部解剖結構，同時降低重建過程中產生的偽影。透過有效感受野（ERF）與注意力圖（attention map）進行可解釋性分析後發現，INet能有效利用更廣泛的解剖脈絡資訊，並產生具任務適應性的特徵響應，說明其輕量化架構仍能維持良好效能的原因。

為評估生成偽電腦斷層影像之物理保真度，本研究進一步利用真實電腦斷層與偽電腦斷層影像進行穿顱聚焦超音波（tFUS）模擬。由INet生成之偽電腦斷層影像所推導出的聲學模擬參數，與真實電腦斷層所得到之結果未呈現統計上的顯著差異，顯示INet能保留可靠超音波傳播模型所需之顱骨特性。此外使用INet偽電腦斷層進行模擬所得到的結果，與真實電腦斷層模擬結果具有相當的一致性，並且在大部分評估之模擬指標中皆較UNet偽電腦斷層表現出更佳的結果。

綜合以上結果，INet在影像重建準確度、計算效率以及模型可解釋性之間取得良好的平衡。本研究所提出的方法能以輕量化架構生成偽電腦斷層影像，同時保留進行下游穿顱聚焦超音波模擬所需之結構與物理特性，顯示其具備應用於僅使用磁振造影影像之超音波治療規劃流程的潛力。

Pseudo-CT generation from MR images provides a potential solution for reducing the reliance on CT acquisition while preserving the physical information required for therapeutic applications. However, existing pseudo-CT generation networks often rely on complex architectures with high computational costs, limiting their practical deployment. In this study, we propose INet, an ultra-lightweight and explainable network for pseudo-CT generation.

Compared with the conventional UNet, INet reduces computational complexity to 20.5% in FLOPs and 0.58% in parameter count while maintaining comparable reconstruction performance. On the independent testing set, INet achieved statistically comparable performance to UNet and UNet3+ across multiple quantitative metrics, and significantly outperformed UNeXt. Qualitative analysis further demonstrated that INet better preserved skull continuity, cortical boundaries, and local anatomical structures while reducing reconstruction artifacts. Explainability analysis using ERF and attention maps revealed that INet effectively utilizes broader anatomical context and generates task-adaptive feature responses, providing insights into how the lightweight architecture maintains competitive performance.

To evaluate the physical fidelity of the generated pseudo-CT images, tFUS simulation was performed using real CT and pseudo-CT images. The acoustic simulation parameters derived from INet-generated pseudo-CT images showed no statistically significant differences compared with those obtained from real CT images, indicating that INet preserves the skull characteristics required for reliable ultrasound propagation modeling. Furthermore, simulations using INet-generated pseudo-CT images achieved comparable agreement with real CT-based simulations and consistently demonstrated improved performance compared with UNet-generated pseudo-CT images across most evaluated simulation metrics.

These results demonstrate that INet provides an effective balance between reconstruction accuracy, computational efficiency, and interpretability. The proposed framework enables lightweight pseudo-CT generation while preserving the structural and physical properties required for downstream tFUS simulation, supporting its potential application in MR-only ultrasound treatment planning workflows.