DC 欄位 |
值 |
語言 |
DC.contributor | 機械工程學系 | zh_TW |
DC.creator | 語嫣 | zh_TW |
DC.creator | Nazish Murad | en_US |
dc.date.accessioned | 2023-11-8T07:39:07Z | |
dc.date.available | 2023-11-8T07:39:07Z | |
dc.date.issued | 2023 | |
dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=107383610 | |
dc.contributor.department | 機械工程學系 | zh_TW |
DC.description | 國立中央大學 | zh_TW |
DC.description | National Central University | en_US |
dc.description.abstract | 本論文基於近红外光 (near-infrared, NIR)及擴散光學造影 (Diffuse Optical Imaging, DOI)技術,結合不同特性學習層發展深度學習網路以重建乳腺腫瘤識別及定位的擴散光學斷層造影 (Diffuse Optical Tomography, DOT)。為克服有限數據帶來的挑戰,研究生成具有多個仿體和不同部位置入物的大型數據集,並設計一單一深度神經網路 (Deep Neural Network, DNN)重建吸收集、散射係數。透過大量的實驗與優化,決定所提出的網路架構及各層的順序與排列方式,該網路可學習分層象徵並透過卷積及特定順序排列的批次歸一化層提取持續變化的複雜特徵。所提出的 DNN模型使用不同的仿體數據集來進行評估,而提出的Tikhonov正則化 (Tikhonov Regularization, TR) 方法和其他人工神经網路 (Artificial Neural Network, ANN)相比 (如倒傳遞神經網路, U-net、Dense-net 等)表現出更加優越的性能。此外,包含批次歸一化的DNN 模型可提高空間分辨率,提供合適的對比度尺寸细節(Contrast-and-Size Detail, CSD)分析。
一種端到端的靈活深度學習框架被用於檢測早期乳腺癌並重建其光學特性。所提出的Periodic-net框架獲得反算模型中的不均勻精確重建與評估。Periodic-net 允許有效組合不同空間尺度的濾波器(filter),允許網路同時提取細粒度的細節與更全局的上下文訊息。透過組合這些濾波器,Periodic-net可有效捕捉多樣特徵而不犧牲計算效能 (Dense-net:31s, U-net:20s, Periodic-net:3s)。本研究也採用了複合式深度學習模型(hybrid deep learning model)用於擴散光學成像中的腫瘤檢測。來自前向模型的訊號和反向計算的圖像在單一解碼器中結合,彌補了直接及後處理方法之間的差距。複合式深度學習模型達到了最先進的重建精度 (模擬數據集的SSIM:0.88 和 PSNR:33.50;量測數據集的SSIM:0.52 和 PSNR:31.74)。
此外,本研究使用了遷移學習方法解決圓形仿體的侷限並提高重建方法的多功能性。遷移學習方法用來調整先前開發的Periodic-net 網路架構來處理橢圓仿體,利用從圓形仿體數據集所得到的先驗知識與預訓練權重,透過新收集的橢圓形仿體數據對網路進行微調。
總結來說,本文提出了用於DOI重建的創新深度學習架構,克服了有限數據集、空間分辨率提升、精確腫瘤檢測及光學特性高效重建等多項挑戰。這些進步有效推進DOI技術的演進,並有望增强乳腺腫瘤早期檢测和定性的齡床應用。與最先進的競爭對手相比,重建出來的圖像顯示. | zh_TW |
dc.description.abstract | This thesis focuses on the development of deep learning networks by combining properties of various layers for the reconstruction of diffuse optical imaging (DOI) in the context of near-infrared (NIR) diffuse optical tomography (DOT) for breast tumor identification and localization. Large datasets with multiple phantoms and inclusions at various positions are generated to overcome the limited dataset challenge. Additionally, a single deep neural network (DNN) model is designed to reconstruct both absorption and scattering coefficients. Extensive experimentation and optimization determined the specific order and arrangement of layers in proposed networks. The network can learn hierarchical representations and extract increasingly complex features by applying convolutional layers followed by batch-normalization layers arranged in a specific order. The proposed DNN models are evaluated using a varied phantom dataset, demonstrating superior performance compared to the Tikhonov Regularization (TR) method and other artificial neural networks (ANN), i.e., backpropagation neural network, U-net, Dense-net. Moreover, the inclusion of a batch normalization layer in the DNN model results in improved spatial resolution, providing a suitable Contrast-and-Size Detail (CSD) analysis.
An end-to-end flexible deep learning framework is developed for DOI to detect breast cancer and reconstruct its optical properties in the early stages. The proposed Periodic-net algorithm achieves accurate reconstruction and evaluation of inhomogeneities in an inverse model. Periodic-net allows for the efficient combination of filters at different spatial scales, allowing the network to extract fine-grained details and more global context information simultaneously. By combining these filters, the Periodic-net efficiently captures a wide range of features without sacrificing computational efficiency (Dense-net: 31s, U-net: 20s, Periodic-net: 3s). This study also incorporated a hybrid deep learning model for tumor detection in diffuse optical imaging. Signals from the forward model and images from the inverse computations are combined in a single decoder, bridging the gap between direct processing and post-processing methods. The hybrid network achieves state-of-the-art reconstruction accuracy (SSIM: 0.88 and PSNR: 33.50 for simulation and SSIM: 0.52 and PSNR: 31.74 for measure dataset).
Additionally, in order to address the limitations of circular phantoms and improve the versatility of the reconstruction approach, transfer learning is utilized in this research. Transfer learning is employed to adapt the previously developed Periodic-net architecture to handle elliptical phantoms. The network is fine-tuned using the newly acquired elliptical phantom dataset by leveraging the knowledge and pre-trained weights obtained from the circular phantom dataset.
Overall, this thesis presents innovative deep-learning networks for diffuse optical imaging reconstruction, overcoming challenges related to limited datasets, improved spatial resolution, accurate tumor detection, and efficient reconstruction of optical properties. These advancements contribute to diffuse optical imaging and hold promise for enhanced clinical applications in the early detection and characterization of breast tumors. The reconstructed images demonstrate higher quality, increased small detail preservation, improved noise immunity, sharper edges, and reduced artifacts compared to state-of-the-art competitors. | en_US |
DC.subject | 擴散光學斷層造影 | zh_TW |
DC.subject | 組織光學特性 | zh_TW |
DC.subject | 批次歸一化 | zh_TW |
DC.subject | Periodic-net | zh_TW |
DC.subject | 複合式模型 | zh_TW |
DC.subject | 遷移學習 | zh_TW |
DC.subject | 深度卷積網路 | zh_TW |
DC.subject | Tikhonov正則化 | zh_TW |
DC.subject | 域轉換 | zh_TW |
DC.subject | 逆問题 | zh_TW |
DC.subject | Diffuse Optical Imaging | en_US |
DC.subject | tissue optical properties | en_US |
DC.subject | Batch-normalization network | en_US |
DC.subject | Periodic-Net | en_US |
DC.subject | Hybrid Network | en_US |
DC.subject | transfer learning | en_US |
DC.subject | deep convolutional neural networks | en_US |
DC.subject | Tikhonov regularization | en_US |
DC.subject | domain transformation | en_US |
DC.subject | inverse problem | en_US |
DC.title | Enhancing Diffuse Optical Imaging: Utilizing Deep Learning Networks for Accuracy and Efficiency Improvement in Image Reconstruction of Optical Properties | en_US |
dc.language.iso | en_US | en_US |
DC.type | 博碩士論文 | zh_TW |
DC.type | thesis | en_US |
DC.publisher | National Central University | en_US |