增進基於核磁共振之影像體學特徵的前處理以提高影像體學預後預測模型之準確度以及泛用性

Abstract

腫瘤治療預後評估是癌症醫學的基石，放射影像則扮演癌症治療評估的關鍵角色，腦部轉移腫瘤接受單獨立體定位放射手術後，相對於傳統的全腦放射線治療，雖然會有較好的認知功能及生活品質，術後但顱內遠端復發腫瘤的比例較高，且這些個案通常伴隨著較差的預後，過去曾有研究以腦內轉移腫瘤體積、顱內轉移腫瘤顆數、全身腫瘤惡化與否來進行預後預測，但預測效果並沒有良好的泛用性及準確性，因此發展一個精準而穩固的轉移腦部腫瘤遠端顱內復發預測模型是臨床上一個重要的課題。影像體學是基於分析醫學影像的影像體特徵是一個新穎、可行且可擷取巨量特徵的方法，標準化影像體特徵結取方法，包括經由引用深度學習影像擴增方法中的平移、旋轉及脹縮，來篩選得到具穩固性的影像體特徵，並加以結合基於機器學習的非線性的降維演算法，希望可以並提升其泛用性，並保有預後預測效果，本研究目標在建立具泛用性且精確性的影像體學預後預測模型，以期能幫助選擇最適合腦轉移腫瘤的第一線治療方式。本計畫預計分析接受立體定位放射手術治療之腦轉移腫瘤病人之磁振造影影像資料集，嘗試預測腦轉移腫瘤受立體定位放射手術之腫瘤復發預後，並建立影像研究之標準流程。

Early tumor prognosis prediction is the cornerstone of cancer medicine. Radiographic imaging plays an important role to describe and predict the outcome of cancer treatment. Whole-brain radiotherapy (WBRT) possessed the historical essential role for multiple brain metastases (BM) treatment in past decades, whereas emerging evidence by several randomized trials supported stereotactic radiosurgery (SRS) alone, instead of upfront WBRT, to be the preferred treatment modality for limited numbers of BM (defined as 1 to 4). SRS alone provides better cognitive and quality of life (QOL) outcomes compared with WBRT and does not compromise with a detrimental effect on survival. However, time and labor costs to design an SRS plan are substantially higher than WBRT, and increased risk of intracranial progression after SRS alone also requires additional salvage SRS or WBRT. Historically, some studies provided a predictive model incorporating pre-SRS tumor volume, tumor numbers, and systemic cancer progression status to predict the intracranial distant progression, but those models lack accuracy and generalization. Radiomic analysis based on radiographic imaging is a novel and efficient approach to extract extensive image features followed by reducing the dimensionality of the features to robust key components with the assistance of high-throughput computing. In addition, radiomic feature extraction by using image augmentation tips, including translation, rotation, and inflation/retraction, and radiomic features selection by using infraclass variation coefficient will provide a robust way to get radiomic features. This study aims to use a brain Magnetic Resonance Imaging (MRI) dataset derived from the National Taiwan university hospital radiosurgery imaging database to develop a machine-learning-based robust radiomic model to predict the distant intracranial progression of patient of BMs treated with SRS. Furthermore, we also establish a standard workflow to analyze the radiographic imaging biomarker.