利用深度卷積神經網路辦識超音波影像增強診斷分化型甲狀腺癌

Taipei Medical University Institutional Repository > 醫學院 > 人工智慧醫療碩士在職專班 > 博碩士論文 > Item 987654321/61667

請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/61667

題名:	利用深度卷積神經網路辦識超音波影像增強診斷分化型甲狀腺癌 Using Deep Convolutional Neural Networks for Enhanced Ultrasonographic Image Diagnosis of Differentiated Thyroid Cancer
作者:	陳維健 CHAN, WAI KIN
貢獻者:	醫學院人工智慧醫療碩士在職專班彭徐鈞
關鍵詞:	甲狀腺癌 Thyroid cancer
日期:	2021-06-16
上傳時間:	2022-04-28 23:23:28 (UTC+8)
摘要:	背景臨床上甲狀腺癌大部分為濾泡上皮細胞來源分化型甲狀腺癌。雖然最常見為乳突癌(PTC)，但仍有幾種相對少見但較難診斷的病理分類如濾泡癌(FTC)等。故希望利用遷移學習方式訓練深度卷積神經網路(deep convolutional neural network)，增強分化型甲狀腺癌的診斷能力。研究材料與方法本研究回溯財團法人長庚紀念醫院2003年1月至2020年7月收集的病歷等資料進行分析。共收納421位DTC病例及391位良性甲狀腺腫瘤的病例。以監督式學習(supervised learning)及遷移學習(transfer learning) 的方式訓練InceptionV3、ResNet101及VGG19辨識惡性與良性甲狀腺腫瘤。結果病例根據惡性及良性的病理組織學被分為PTC、FTC、濾泡型甲狀腺乳突癌(Follicular variant of PTC)、Hürthle細胞癌(Hürthle Cell Carcinoma)以及良性結節。超音波影像經過裁切及前處理後收集到的腫瘤影像共有2308張。InceptionV3、ResNet101及VGG19經過遷移學習後總準確率分別達76.5%、77.6%及76.1%。InceptionV3及VGG19分別有較高的靈敏度(Sensitivity)及特異性(specificity)。Area under curve(AUC)在InceptionV3、ResNet101及VGG19分別達0.82、0.83及0.83。結論使用遷移學習的方式訓練的深度卷積神經網路已能夠應用在臨床上，且可以提升診斷分化型甲狀腺癌的準確率，包括診斷最困難的甲狀腺濾泡癌等。在未來人工智慧電腦醫療診斷輔助工具(computer aided diagnosis)配合臨床醫師能簡單利用超音波準確地診斷甲狀腺癌。 Background Most thyroid cancers are differentiated thyroid cancer derived from follicular epithelial cells. Although the most common is papillary thyroid carcinoma (PTC), there are still several relatively rare but difficult to diagnose pathological classifications such as follicular thyroid carcinoma(FTC), etc. In the era of artificial intelligence, deep convolutional neural network is believed to enhance our clinical diagnostic ability of differentiated thyroid cancer. Materials and methods This study retrospectively analyzed the medical records and data collected at Chang Gung Memorial Hospital from January 2003 to July 2020. A total of 421 DTC cases and 391 benign thyroid tumor cases were enrolled. InceptionV3, ResNet101 and VGG19 as well-known pre-trained CNNs with high accuracy, were then re-trained by supervised learning and transfer learning method to identify malignant and benign thyroid tumors. Result The cases enrolled were classified into PTC, FTC, Follicular variant of PTC, Hürthle Cell Carcinoma and Benign group based on malignant and benign histopathology. The collected ultrasound images were cropped and pre-processed reaching a total of 2308 tumor images. The total accuracy of InceptionV3, ResNet101 and VGG19 after transfer learning reached 76.5%, 77.6% and 76.1%, respectively. InceptionV3 and VGG19 have the highest sensitivity and specificity respectively. Area under curve (AUC) reaches 0.82, 0.83 and 0.83 in InceptionV3, ResNet101 and VGG19 respectively. Conclusion Deep convolutional neural networks re-trained by transferred learning method can be deployed in clinical scenarios and being able to increase diagnostic accuracy in most differentiated thyroid cancers including follicular cancers. In the near future, thyroid cancers can be diagnosed easier and more accurately when combining computer aided diagnosis and professional clinicians.
描述:	碩士指導教授：彭徐鈞委員：劉文德委員：崔博翔
資料類型:	thesis
顯示於類別:	[人工智慧醫療碩士在職專班] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	138	檢視/開啟

檢視Licence

在TMUIR中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....