器官和支氣管疾病
頂刊論文
1. How Far Are We from Using Radiomics Assessment of Gliomas in Clinical Practice?
Radiomics從數(shù)據(jù)中挖掘有臨床價(jià)值的信息屯烦。簡(jiǎn)而言之，包含圖像采集、分割寞酿、特征提取和特征選擇這4個(gè)必要的過(guò)程趟妥。通常提取到的特征有四類：
- 一階統(tǒng)計(jì)信息：mean, variance, kurtosis, skewness and 描述整個(gè)腫瘤直方圖分布的entropy
- 紋理特征：homogeneity， contrast疤坝，gray-level nonuniformity, cluster tendency, and harder-to-picture features such as short run emphasis
- 小波特征：features dependent on spatial frequencies
- 形狀特征：volume, surface area, sphericity, compactness, and ?atness
前三類特征是“agnostic” features兆解，他們是數(shù)學(xué)上定義的量化描述子，還沒(méi)有被收錄到放射學(xué)詞典中跑揉。另一方面形狀特征通常指“semantic” features锅睛，他們中的一些已經(jīng)被收錄到放射性詞典中埠巨。
特征的篩選和模型的選擇也是需要重點(diǎn)考慮的。
一個(gè)典型的例子：Radiomic MRI Phenotyping of Glioblastoma: Improving Survival Prediction 使用variable-hunting feature selection將796個(gè)特征whittle為18個(gè)现拒，采用RSF(random survival forest)預(yù)測(cè)辣垒。
類似這樣的研究raise如下重要的問(wèn)題：
- 這些特征跟我們已經(jīng)熟知的疾病生物學(xué)特征有什么關(guān)聯(lián)以及如何關(guān)聯(lián)？
這些特征能增加我們對(duì)疾病的理解嗎印蔬？ - 有沒(méi)有改變或增加我們對(duì)疾病已經(jīng)有的認(rèn)知勋桶？
Some work has begun in this arena: First-order features obtained from histogram-based methods have been shown to relate to tumor cellularity (5). Textural features, on the other hand, re?ect tumor heterogeneity. Tey are potential markers for tumor aggressiveness and, possibly, response to therapy (2,6). Tumor shape features (the only “semantic” features) have also been shown to relate to tumor aggressiveness (2).
盡管對(duì)radiomic features的知識(shí)越來(lái)越多，但是困難仍然存在：如何整合這些結(jié)果到臨床實(shí)踐中侥猬。 （實(shí)驗(yàn)結(jié)果的可重復(fù)性問(wèn)題以及機(jī)器學(xué)習(xí)的黑箱問(wèn)題）
