| 摘要: | 甲狀腺腫瘤是內分泌系統常見的病症。甲狀腺腫瘤的危害不大，不影響日常生活。不過當甲狀腺腫瘤被診斷為惡性腫瘤，或是腫瘤過大壓迫到氣管，就需要即時手術治療。通常使用超音波檢測甲狀腺腫瘤，這是一個簡單、快速和省錢的方法，也是非侵入性和無輻射的檢測方法。不過為了近一步的診斷甲狀腺腫瘤的良惡性，還是需要使用細微針頭提取腫瘤細胞，再用顯微鏡診斷腫瘤類型。目前判讀超音波影像中的甲狀腺腫瘤，還是依賴於臨床醫師的肉眼，這種判定方式不僅耗時且會因經驗而影響判讀。
本研究提出一個基於YOLO v4深度神經網路開發一個自動檢測超音波甲狀腺腫瘤之檢測器，實現快速偵測甲狀腺腫瘤。我們選擇七個預訓練的卷積神經網路作為特徵提取網路，結合YOLO v4網路，創建了七個自定義的YOLO v4檢測器。我們使用影像修復法移除了超音波影像中的記號，然後使用數據增量增加影像數量和提高影像對比度。經過5-fold交叉驗證後，當特徵提取網路是NASNet-Large時，模型在測試集的評估指標達到:平均Precision為92.2%、平均Recall為 85.7%、平均F1-score為 88.8%和平均Average Precision為 84%。最後我們設計一個圖形使用者介面，輔助臨床醫師方便快速診斷超音波影像上的甲狀腺腫瘤。
;The thyroid tumor is a common condition in the endocrine system. Thyroid tumors typically pose minimal risk and do not significantly impact daily life. However, when a thyroid tumors is diagnosed as a malignant tumor or when the tumor is excessively large and compressing the trachea, immediate surgical treatment is necessary. Thyroid tumors are typically detected using ultrasound, a simple, quick, cost-effective, non-invasive, and radiation-free method of examination. However, for a more detailed assessment of the benign or malignant nature of thyroid nodules, fine-needle aspiration (FNA) is required to extract cells from the tumor for microscopic examination to determine the type of tumor. Currently, the interpretation of thyroid nodules in ultrasound images still relies on the visual assessment of clinical physicians. This method is not only time-consuming but also susceptible to interpretation variations based on experience, affecting the accuracy of diagnosis.
This study proposes the development of an automatic thyroid nodule detection system based on the YOLO v4 deep neural network. This system aims to achieve rapid detection of thyroid nodules in ultrasound images. We selected seven pre-trained convolutional neural networks as feature extraction networks and combined them with the YOLO v4 network, creating seven customized YOLO v4 detectors. We utilized image inpainting techniques to remove artifacts from the ultrasound images and employed data augmentation to increase the image quantity and enhance image contrast. After 5-fold cross-validation, when the feature extraction network was NASNet-Large, the model achieved the following evaluation metrics on the test set: average Precision = 92.2%, average Recall = 85.7%, average F1-score = 88.8%, and average AP = 84%. We concluded by designing a graphical user interface to assist clinical physicians in efficiently diagnosing thyroid nodules in ultrasound images, providing them with convenience and speed during the diagnostic process. |
