超音波彈性影像技術（Elastography）能直觀呈現生物組織的彈性特性，廣泛應用於臨床診斷。然而，傳統彈性影像設備價格昂貴且成像品質易受操作者技術差異及施力不均等因素影響，限制了診斷的客觀性與一致性。有鑑於此，本研究基於深度學習的雙階段彈性影像生成網路（Dual Stage Elastography Generation Network, DSEG-Net），透過一般臨床常見之B-mode超音波影像，生成脛部（SHANK）部位拔罐療法前後之高品質灰階彈性影像，有效降低對專業設備的依賴，並改善診斷的主觀性。 不同以往多針對乳房或甲狀腺的彩色彈性影像分析，本研究首次聚焦於脛部 部位的灰階彈性影像處理，以拔罐療法前後的影像生成任務提高模型泛化性。在臨床上可藉由拔罐前後彈性影像的視覺化比較，直觀評估肌肉的緊繃或放鬆程度，協助判定拔罐療法之效果及療效追蹤。共從24名患者收集857張超音波彈性影像資料，並透過 DSEG-Net 進行雙階段生成，藉此有效量化肌肉鬆軟或僵硬程度的改變。在此過程中引入跨注意力機制（Cross Attention），提升生成影像的視覺品質與物理意義。 經視覺比較與量化評估，本研究之方法在結構相似性指標（SSIM）、影像生成真實性指標（FID）以及評估模型（Evaluation Model）判斷生成彈性影像是否準確呈現肌肉彈性特徵的能力，均優於現有方法，證實生成影像的高度臨床實用價值與品質穩定性，展現未來廣泛臨床應用之潛力。;Ultrasound elastography provides intuitive visualization of tissue elasticity and is widely used in clinical diagnosis. However, traditional elastography equipment is expensive and sensitive to operator technique, limiting consistency and accessibility. To address this, we propose a deep-learning-based Dual Stage Elastography Generation Network (DSEG-Net) that generates high-quality grayscale elastography images of the shank before and after cupping therapy, using only standard B-mode ultrasound images. This approach reduces reliance on specialized devices and improves diagnostic objectivity. Unlike previous studies that focused on color elastography of the breast or thyroid, our work is the first to target grayscale elastography of the shank, enhancing model generalization. Clinically, comparing elastography images before and after cupping can help visualize muscle stiffness or softness, supporting treatment assessment and tracking. A total of 857 images from 24 patients were collected, and a cross attention mechanism was introduced to improve both visual quality and physical relevance. Through visual and quantitative evaluation, DSEG-Net outperformed existing methods in SSIM, FID, and the evaluation model’s ability to verify whether the generated images accurately reflect muscle elasticity. The results confirm the method’s potential for reliable and practical clinical use.
