| DC 欄位 |
值 |
語言 |
| DC.contributor | 跨領域轉譯醫學研究所 | zh_TW |
| DC.creator | 黃信哲 | zh_TW |
| DC.creator | Oscar Huang | en_US |
| dc.date.accessioned | 2025-1-22T07:39:07Z | |
| dc.date.available | 2025-1-22T07:39:07Z | |
| dc.date.issued | 2025 | |
| dc.identifier.uri | http://ir.lib.ncu.edu.tw:444/thesis/view_etd.asp?URN=105888003 | |
| dc.contributor.department | 跨領域轉譯醫學研究所 | zh_TW |
| DC.description | 國立中央大學 | zh_TW |
| DC.description | National Central University | en_US |
| dc.description.abstract | 在本篇論文中，我們根據量子化學在研究高分子材料的理論基礎—密度泛函理論(Density Functional Theory)—上，開發了一種可討論無窮維度資料架構的半無監督(Semi-unsupervised) 的機械學習演算法。由於該方法被我們架構在微分幾何的數學架構之下， 並擁有深度學習當中的變分自編碼器 (Variational AutoEncoder) 的潛變空間 (Latent Space) 架構，因此我們稱之為幾何快速資料密度泛函轉換(Geometric Fast Data Density Functional Transform, g-fDDFT)。在理論印證的前期，我們使用 g-fDDFT 來協助辨識人體的動作模式。我們首先將動作的時序訊號映射到特定的偽物理空間中，先利用能量穩定理 (Energy Stability Principles) 決定最可能的資料群數 (Clustering Numbers)，再透過尋找局部能量最低值 (Local Energy Minimum) 來確定對應於每個群(亦即不同動作狀態)的資料邊界 (Cluster Boundaries)，其中資料的狀態移動方向和狀態間的過度區域即成為重要的動作辨識特徵。而不同於傳統方法，我們僅以一個三軸重力感測器收集資訊，透過結合動作特徵與感測器架構作為先驗資訊，實驗結果證實最可能的動作狀態可分類為日常生活中的四種常見人體姿勢，且動作誤差和雜訊不會對結果產生明顯影響。我們也將此方法用於模擬翻身情境，結果證實應用於老人和嬰兒翻身監控上之潛力。此外，我們進一步最佳化 g-fDDFT，並將印證實例轉至討論灰階的醫學影像矩陣同構映射至能量空間的方法，期待能藉此實現以非監督式方法辨識腦腫瘤的病灶形態。在不斷最佳化 g-fDDFT 理論架構時，我們引入了幾何深度學習 (geometric deep learning) 的概念架構和圖神經網路 (graph neural network, GNN) 的度規 (metrics) 方法。這兩者的理論整合，實踐了 g- fDDFT 的網格化密度泛函 (gridized density functional) 簡化方法，並建立了非監
督式的圖形特徵感知機制。而在最佳化的數學架構上，g-fDDFT 依然使用全域 卷積核 (Global Convolutional Kernels)來辨識與提取最可能的醫學影像的病灶邊界並加以分割。在 g-fDDFT 的模型架構中，我們所設計的變分自編碼器輔助模組(Variational AutoEncoder-assisted Module)能將原始卷積的計算複雜度從 ??(??3)降低至 ??(?? log ??)，從而顯著加快全域卷積操作的速度。在理論印證時，我們使用各種公開資料集驗證模型的性能並討論其限制。在超大型 3D 資料集如 MACCAI 國際研討會中所提供的 BraTS 系列競賽的公開 MRI 腦腫瘤醫學影像資料集，每個病例的平均模型推論時間 (Inference Time) 約為 1.76 秒。我們所設計的網格化密度泛函具有活化能力 (Activation Capability)， 且能與梯度上升 (Gradient Ascent) 操作產生交互作用並可用於分析操作產生的協同能力，因此可以模組化併入深度神經網路的自動化學習流程 (Pipeline) 中。理論架構上，我們也開發了幾何穩定性 (Geometric Stability) 與相似性收斂 (Similarity Convergence)之演算法來提高對病灶影像進行非監督式辨識和分割的準確性，並能達到傳統醫學影像深度神經網路之要求標準，我們的模型 Dice 分數的中位數高於 0.75。我們的實驗顯示，g-fDDFT 若與一個簡單神經網路協同作用，能使模型訓練時間(Training Time)和推論時間分別加快了 58％ 和 51％，Dice 分數則提升至 0.9415。g-fDDFT 的這項優勢能促進跨學科應用和臨床研究中的快速計算建模。 | zh_TW |
| dc.description.abstract | We established an innovative approach for human motion recognition with the geometric fast data density functional transform (g-fDDFT). We mapped the temporal motion signals into a specific physical space and used energy stability principles to determine the most probable number of clusters. The clusters corresponded to different motion states, and their boundaries were identified by locating local energy minima. The direction of state migration and the transition regions between these states became key motion features. Unlike traditional methods, we used only a single tri-axial gravitational sensor for data collection. By integrating motion features and the sensor architecture as prior information, experimental results showed that the most probable motion states can be classified into four common human postures observed in daily life, with error motions and noise having minimal impact on the results. This method was also successfully applied in simulating turning-over scenarios, confirming its potential for monitoring the movements of elderly individuals and infants. Additionally, we optimized the theoretical framework of g-fDDFT, which isomorphically mapped gray- level medical image matrices onto energy spaces, enabling the unsupervised recognition of lesion morphology. By incorporating geometric deep learning (GDL) architecture and metrics from graph neural networks (GNN), the gridized density functionals of g-fDDFT established an unsupervised feature-aware mechanism. This mechanism uses global convolutional kernels to extract the most probable lesion boundaries and perform segmentation. The computational complexity was reduced from ??(??3) to ??(?? log ??) by the AutoEncoder-assisted module, significantly speeding up the global convolution operations. We validated the model′s performance using various open datasets and discussed its limitations. In large 3D datasets, the
average inference time for each case was 1.76 seconds. The gridized density functionals demonstrated activation capabilities that synergize with gradient ascent operations, allowing for modularization and embedding into the pipelines of deep neural networks. Algorithms for geometric stability and similarity convergence further enhanced unsupervised lesion recognition and segmentation accuracy, achieving performance standards comparable to conventional deep neural networks, with a median Dice score exceeding 0.75. Our experiments revealed that the synergy between fDDFT and a simple neural network improved training and inference times by 58% and 51%, respectively, and raised the Dice score to 0.9415. This enhancement facilitates rapid computational modeling in interdisciplinary applications and clinical research. | en_US |
| DC.subject | 半無監督的機械學習演算法 | zh_TW |
| DC.subject | 無監督的機械學習演算法 | zh_TW |
| DC.subject | MRI 腦腫瘤醫學影像辨識 | zh_TW |
| DC.subject | 快速推論時間 | zh_TW |
| DC.subject | 深度神經網路的自動化學習流程 | zh_TW |
| DC.subject | 自編碼器輔助模組 | zh_TW |
| DC.subject | Semi-unsupervised Machine learning algorithms | en_US |
| DC.subject | unsupervised Machine learning algorithms | en_US |
| DC.subject | MRI Brain Tumor Medical Image Identification | en_US |
| DC.subject | fast Inference Time | en_US |
| DC.subject | Automated learning for deep neural networks Pipeline | en_US |
| DC.subject | Variational AutoEncoder-assisted Module | en_US |
| DC.title | 幾何快速資料密度泛函轉換之理論開發以及於三維的人體姿勢辨識與追蹤、高維腫瘤影像病灶辨識與分割之研究 | zh_TW |
| dc.language.iso | zh-TW | zh-TW |
| DC.type | 博碩士論文 | zh_TW |
| DC.type | thesis | en_US |
| DC.publisher | National Central University | en_US |