生成模型於MRI膝關節影像之應用

Abstract

退化性關節炎(OA)是一種常見影響膝關節的疾病，可能導致病患的疼痛、嚴重的行動不便，甚至嚴重的OA有時需要進行關節置換手術。OA造成疼痛的成因複雜，可能源自於膝關節的腫脹、骨質弱化和軟骨磨損等變化。並且疼痛受個體感知因素影響，個體對疼痛的感知並不相同，使得臨床上難以將個體回報的疼痛指數與影像連結起來。因此，透過深度學習模型將醫學影像與OA疼痛連結具有臨床重要性，其中磁振造影 (Magnetic Resonance Imaging, MRI)膝關節影像能提供最全面的資訊。基於上述原因，本研究利用生成式模型針對MRI影像進行兩個不同方向的OA研究。 第一個研究提出兩階段生成對抗網路(Generative Adversarial Networks, GAN)針對OA與軟骨磨損的關係進行研究，透過生成模型合成OA患者軟骨未磨損前的MRI影像進一步討同個OA病患軟骨磨損前後的差異。實驗結果顯示，正常股骨軟骨平均體積與脛骨軟骨平均體積分別較OA患者大3.3%及4.5%，股骨軟骨平均包覆股骨長度與脛骨軟骨平均包覆脛骨長度較OA患者長1.2%及3%。 第二個研究提出兩階段對比式生成對抗網路 (Contrastive GAN) 及生成擴散模型 (Diffusion Model) 探討骨內病變 (Bone Marrow Lesions, BML) 與積水(effusion) 對OA病患造成疼痛貢獻，並量化此貢獻。 第一階段的對比式生成對抗網路負責篩選出所有可能造成疼痛貢獻之區域，同時模型存在一投影層 (projection layer) 透過對比式學習，模型學習度量OA病患的疼痛程度。第二階段的生成擴散模型負責學習消除MRI膝關節影像指定區域的病變。其中，為了解決二維模型缺乏三維生理結構連續性，例如: 肌肉、骨骼及病變，的生成能力，以及解決三維生成模型過於龐大，耗費計算資源的問題，提出具創新性的雙編碼器生成擴散模型。透過提出的兩階段式生成模型，獨立控制生成消除單一或多個病變的MRI影像模擬病患病變修復後的情況，並經由第一階段對比式生成對抗網路的投影層藉由均勻流形逼近和投影 (Uniform Manifold Approximation and Projection, UMAP) 技術量化病變對OA病患的疼痛貢獻。 實驗結果顯示，提出的兩階段模型在影像所有採用的生成品質度量上包含Fréchet Inception Distance (FID), Inception Score (IS), Structural Similarity Index Measure (SSIM), Multi-Scale Structural Similarity Index Measure (MS-SSIM), Learned Perceptual Image Patch Similarity (LPIPS) 以及Flow difference-weighted LPIPS (FloLPIPS) 上超越2維最先進 (state-of-the-art, SOTA) 的模型及三維最先進的模型。值得一提的是，透過UMAP量化單獨消除BML或effusion的任務上顯示，消除effusion對疼痛減緩的貢獻較為顯著，未來可以成為臨床上治療的依據。

Osteoarthritis (OA) is a prevalent condition impacting knee joints, often leading to significant pain and mobility restrictions. In severe cases, it may necessitate joint replacement surgery. The origins of OA pain are complex, involving knee joint changes such as swelling, bone weakening, and cartilage loss. Since pain perception varies individually, linking patient-reported pain with imaging findings is challenging. Thus, employing deep learning models to correlate medical imaging with OA pain is clinically significant, particularly MRI providing comprehensive knee joint information. Given these considerations, this study employs generative models in MRI to undertake two distinct research directions pertaining to OA. The first research involves a two-stage Generative Adversarial Network (GAN) to explore the relationship between OA and cartilage erosion. By generating MRI images of OA patients prior to cartilage erosion, the study quantifies differences in cartilage conditions. Results show average femur and tibia cartilage volumes in normal conditions are 3.3% and 4.5% larger, respectively, compared to those with OA. Additionally, the average coverage lengths of femur and tibia cartilages are 1.2% and 3%, respectively, longer in normal conditions. The second research proposes a two-stage Contrastive GAN and a Diffusion Model to explore and quantify the contributions of Bone Marrow Lesions (BML) and effusion to OA pain. The first stage involves a Contrastive GAN that selects regions potentially contributing to pain, with a projection layer utilizing contrastive learning to measure pain severity in OA patients. The second stage, the diffusion model, learns to eliminate specific lesions from MRI knee images. To address the challenge of generating 3D anatomical continuity in 2D models, such as muscles, bones, and lesions, and to overcome the computational demands of 3D models, an innovative dual-encoder diffusion model is introduced. This two-stage model independently controls the generation of MRI images with single or multiple lesion eliminations, simulating and quantifying lesion contributions to OA pain. It employs dimensional reduction techniques, notably Uniform Manifold Approximation and Projection (UMAP). UMAP is a dimensional reduction technique that can find the translation between topological spaces and simplicial sets. The model uses UMAP in conjunction with the projection layer from the first-stage Contrastive GAN. Experimental results show that this two-stage model surpasses both 2D and 3D state-of-the-art models in all employed image quality metrics. Notably, the task of UMAP analysis involving the isolated elimination of BML or effusion indicates that the removal of effusion significantly contributes to pain alleviation, providing a potential basis for future clinical treatment strategies.