根據臨床統計惡性顱內腦瘤平均存活期約僅十五個月左右，兩年內的復發率更是幾乎百分之百，因此於早期進行評估、診斷和治療是非常重要的。同時於醫學影像技術中自動化的腦腫瘤辨識與分割任務的成功與否，對於醫生的診斷協助和患者疾病的診治策略扮演著重要角色。但由於醫生或醫事專家在核磁共振造影影像中手動標記腦瘤位置以及安排手術路徑與計畫需要耗費大量的時間，因此為了提高標記腦瘤位置和輪廓的準確率並同時降低醫生的時間與體力的負擔，本研究透過幾何深度學習模型訓練進行自動腦瘤影像辨識、分割與三維腫瘤體型重構。於研究中我們將使用 BraTS 2020資料集，其中資料集會先經過快速資料密度泛函轉換(fast Data Density Functional Transform)先進行腫瘤特徵強化，並結合能強化三微腦腫瘤特徵的Squeeze-and-excitation 模組於D-Unet (Dimension Fusion U- Net)的編碼器-解碼器架構(Encoder-Decoder Structure)中，以完成一系列的幾何深度學習模型並進行訓練。除了使用 D-Unet，我們也將使用其他各類深度學習模型，如nnU-Net、3D-U-net 等當代最熱門的腦腫瘤分割模型在模型計算複雜度以及Dice分割分數結果進行比較，再將Dice分數最高的模型進行改良以獲取最佳分割結果。我們的研究發現，經過快速資料密度泛函轉換之後的資料集，可以大幅縮短深度學習模型的訓練與推論時間(約在 50%以上)，並且還能提供模型的分割性能。;Clinical statistics reveal an average survival period of approximately fifteen months for malignant intracranial tumors, with an almost 100% recurrence rate within two years. Early assessment, diagnosis, and treatment are crucial in addressing this challenging prognosis. Therefore, the success of automated brain tumor recognition and segmentation tasks in medical imaging technology plays a vital role in assisting physicians with diagnosis and guiding patient treatment strategies. Manual annotation of tumor locations in magnetic resonance imaging (MRI) images and medical professionals planning surgical pathways is time-consuming. To improve the accuracy of tumor localization and contour marking while alleviating the burden on healthcare providers, this study employs geometric deep learning models for automatic brain tumor image recognition, segmentation, and three-dimensional tumor volume reconstruction. In our research, we will use the BraTS 2020 dataset, which first undergoes fast data density functional transformation to enhance tumor features, and combines it with Squeeze-and-excitation, which can enhance the features of three-microbrain tumors. The module is built into the Encoder-Decoder Structure of D-Unet(Dimension Fusion U-Net) to complete and train a series of geometric deep learning models. In addition to using D-Unet, we will also use other types of deep learning models, such as nnU-Net, 3D-U-net and other contemporary most popular brain tumor segmentation models to compare the model computational complexity and Dice segmentation score results, and then the model with the highest Dice score is improved to obtain the best segmentation results. Our research has found that the dataset after fast data density functional transformation can significantly shorten the training and inference time of the deep learning model (about 50% or more), and can also improve the segmentation performance of the model.
