| 摘要: | 腎臟腫瘤在雖不在台灣十大常見癌症之列，但如同歐美國家一樣，近30年來發生率有逐年上升的趨勢。電腦斷層(computed tomography, CT)影像在臨床上是第一線評估腎臟腫瘤的方法之一。如何區分良性和惡性腫瘤在臨床上是很重要的，如此一來可以避免不必要的手術和檢查。然而，目前並沒有很好的方法從電腦斷層影像上區分腎臟的良性和惡性腫瘤。腎臟腫瘤切片也許對於區分良性和惡性腫瘤有幫助，但也伴隨著可能的併發症，譬如腫瘤擴散、出血、?管、假性血管瘤、感染與氣胸等等。如果我們可以從電腦斷層影像上區分良性和惡性腫瘤，這些情況將可以避免。
近年來深度學習被廣泛應用在醫療影像分析上，訓練良好的深度學習模型可以精準的擷取影像特徵進而對影像辨識。因此，我們手動截取電腦斷層上的腎臟腫瘤，以遷移式學習進行訓練，希望可以達到區分腎臟腫瘤類別的效果。
我們收集腎臟腫瘤病人554人，包括67位angiomyolipoma (AML)、34位oncocytoma、246位clear cell renal cell carcinoma (ccRCC)、124位chromophobe renal cell carcinoma (chRCC)以及83位papillary renal cell carcinoma (pRCC)，共4238張電腦斷層影像。將資料較少組別做適當資料增量後，分別使用Inception V3和Resnet 50預訓練模型作分析。Inception V3表現最好的正確率為0.830，Resnet 50最好的正確率為0.849，可見使用深度學習應用在腎臟腫瘤的電腦斷層影像辨識上可以達到不錯的結果。
Renal cell carcinoma (RCC) is the ninth most common cancer in Taiwan, and its’ incidence has been increasing over the past three decades, as well as in Europe and USA. Computed tomography (CT) is one of the first-line imaging method used to evaluate renal masses in clinical practice. Distinguishing these benign renal tumors from malignant renal tumors is clinically important to avoid unnecessary surgical intervention or examination. However, there are no consistently reliable pathognomonic CT scan features that can confidently differentiate benign renal tumors from malignant renal tumors. Renal tumor biopsy (RTB) may be helpful for distinguish benign renal tumors from malignant renal tumors. However, complications associated with RTB include tumor cell seeding along the tract, bleeding, fistula, pseudoaneurysm, infection, and pneumothorax (REF). If confident diagnosis of renal lesions with low or no malignant potential can be achieved from CT images, unnecessary surgeries and diagnostic intervention could be avoided.
Recently, deep learning (DL) methods have been applied to various medical imaging applications. These pretrained Convolutional Neural Networks (CNN) models provide high-quality image features which have been verified in various image classification tasks. In experiments, we compared the performances of the features extracted from hand-crafted features (HCF) to find out which CNN model is suitable for our task
554 patients were enrolled in this study, including 67 patients with angiomyolipoma, 34 patients with oncocytoma, 246 patients with clear cell renal cell carcinoma, 124 patients with chromophobe renal cell carcinoma and 83 patients with papillary renal cell carcinoma, total 4238 computed tomography images. After data augmentation, we used inception V3 and Resnet50 as CNN model for 5-fold training-validation and test. The best accuracy of Inception V3 is 0.8302, and Resnet 50 is 0.8491. Using deep learning to predict renal tumor subtypes could achieve good accuracy |
