使用多參數磁振造影預測外圍區攝護腺癌的格里森分級

Abstract

背景：經醫生臨床診斷病患可能患有攝護腺癌後，需要進一步使用侵入式的病理切片證實，並判定病灶之格里森分級(Gleason grading)，加上考量病患整體身理狀況來決定治療方案。因此使用非侵入的醫學影像輔助系統來預測病理的格里森分級就顯得十分重要，可用於協助放射科醫師尋找病灶及提供治療資訊參考。 方法：本研究目的為使用攝護腺癌患者的磁振造影，預測外圍區攝護腺癌病理上的格里森分級是否高於G4+3。本研究總計使用46名病患資料，39名為訓練數學模型，7名用於獨立測試模型。每名病患使用T2 (T2-weighted images)、ADC (Apparent diffusion coefficient)、DWI (Diffusion-weighted images)三種影像，並經由專業放射科醫師圈選病灶區域，並給定病理切片證實之格里森分級。將病灶區域分割出後，將其切割為數張25*25大小的影像，三種影像各有145張訓練影像和20張測試影像，每張影像可提取360個紋理特徵，三種影像總計可提取每張1080個特徵。透過統計及使用繼承式雙目標基因演算法，配合支援向量機分類器進行最佳化的特徵選取與數學建模，供最佳化預測使用。 結果：提出一組使用7個紋理特徵的最佳化支援向量機，訓練準確度高達100%，測試準確度為80%，靈敏度及特異性分別為66.6%及90.9%。 結論：影像校正、影像擷取、特徵擷取、特徵選取、支援向量機參數設定等步驟皆對預測效能影響重大。本論文提供詳細流程並將特徵選取做到最佳化，因此訓練模型效果良好，部份錯誤之測試影像，推測為該影像已俱備部分嚴重癌症特徵或存在出血干擾等情況進而導致，期望後續藉由增加訓練樣本資料以改進預測效能。

Background: The clinical diagnosis of patients with prostate cancer relies on the physician’s judgment based on pathology texture of invasive biopsy. The decision of treatment options depends on the Gleason grade of biopsy and overall condition of the patient. So it is valuable to predict the histological grade using invasive slices from non-invasive medical images. Methods: This study aims to predict whether the Gleason grading is higher than G4+3 or not from the magnetic resonance imaging (MRI) of peripheral zone prostate cancer to assist radiologists in finding lesions and provide treatment information. In this study, there are 46 patients consisting of 39 patients for training a mathematic model and 7 patients for testing the model. Each patient has three types of images, T2-weighted images, apparent diffusion coefficient, and diffusion-weighted images. A professional radiologist circles the lesion area and determines its Gleason grade for all the MRI images. For every lesion area, a number of subimages with the size of 25 * 25 were extracted. There were 145 training images and 20 test images for each type of images. We extracted 360 texture features from each image and there were 1080 features for a patent with three types of images. A set of informative features were selected by statistical analysis and an inheritable bi-objective combinatorial genetic algorithm with cooperation of support vector machine (SVM). A SVM-based model was established using the identified feature set for prediction of Gleason grade. Results: A set of seven informative features of texture with cooperation of SVM was obtained. The prediction model achieved a training accuracy of 100% and a test accuracy of 80% where the test sensitivity and specificity were 66.6% and 90.9%, respectively. Conclusion: The image calibration, image segmentation, feature extraction, feature selection, and parameter setting of SVM significantly affect the prediction performance. This study provides a detailed procedure of optimal feature selection and achieved a satisfactory result. Some test image with failed prediction reveals that the image has presented some significant cancer features or bleeding conditions resulting in an error judgment. The future work is to increase the number of training patients for advancing prediction performance.