以人工智慧為基礎之影像處理分析而研究結節硬化症之腎臟病灶

Abstract

結節性硬化症 (Tuberous sclerosis complex , TSC) 是一種體染色體顯性遺傳病，每 6,000 至 10,000 名活產嬰兒中就有一人患病，此疾病的致病基因是TSC1和TSC2是負責調控細胞生長的基因，當失去功能時，便無法抑制mTOR，導致細胞生長無法踩煞車，使身體各個器官都會產生腫瘤，而我研究的主題之一就是長在腎臟的腫瘤-腎血管肌脂肪瘤(Renal Angiomyolipoma)。腎血管肌脂肪瘤能會因內源性或外力撞擊破裂，造成嚴重的內出血，這是結節性硬化症主要的死因之一。如果放任腫瘤持續變大，會使腎臟變形，壓迫到其他器官，影響到腎功能。Renal Angiomyolipoma(AML)的治療方法為服用mTOR抑制劑- Everolimus，然而這個藥物若要健保給付有一套規範。Renal Angiomyolipoma是很多型態的腫瘤，要判斷是否合乎用藥標準或觀察用藥效果是件非常耗時的工作，因此希望透過這些研究減輕人力需求。 第一部分研究主題是關於Renal Angiomyolipoma的醫學影像研究。基於醫學影像AI的崛起，可以自動化標出(segmentation)腎臟及腫瘤，有已標記的影像後，希望能用AI的方式自動去比對一個病人前後次的影像 - registraion(對位)，這是一種將兩個影像對齊的技術，透過這個技術，觀察兩次影像的差異。研究有用了幾個對位的方法包括Advanced Normalization Tools(ANTS)，voxelmorph對位，視覺化對位的結果，並且我設計了自動化整個腫瘤判讀流程。希望能做到當醫師輸入一個病人新的影像檔案後，自動化分割，將腫瘤自動分顆，並算出總共有幾顆腫瘤，每個腫瘤的位子和體積大小，並與上次影像做比較，到達節省人力的效果，這是醫學影像部分。 第二部分研究主題是希望能藉由長期收集到的臨床數據用機器學習的方法去預測疾病。台大結節硬化症整合門診迄今已有10年以上的歷史，累積了100多人的臨床數據，這次實驗共有116個病人，共26項臨床特徵，希望能藉由這些數據用機器學習的方法(KNN，SVM，Random Forest，Bagging AdaBoost)來預測病人未來是否有腎出血的風險，和用上述資料看是否能預測病人基因型，雖然NGS技術不像過往那麼高不可攀，但從抽血到有正式的臨床報告也是需要一個月以上的時間，希望藉由機器學習的方式，可以直接藉由臨床的數據及影像推測出病人的基因型，在進一步的用臨床數據去預測未來病人是否有腎出血的可能性。

Tuberous sclerosis complex (TSC) is a chromosomal dominant genetic disease affecting one in every 6,000 to 10,000 live births. The disease is caused by the genes TSC1 and TSC2 responsible for regulating cell growth. When it loses function, mTOR cannot be inhibited, resulting in cell growth unable to step on the brakes, causing tumors in various. One of my research topics is the tumor that grows in the kidney - Renal Angiomyolipoma (AML) . Renal angiomyolipoma can rupture due to endogenous or external impact, resulting in severe internal bleeding, which is one of the main causes of death in tuberous sclerosis complex patients. If the tumor continues to grow, it will deform the kidney, compress other organs, and affect kidney function. The treatment for Renal Angiomyolipoma is to take the mTOR inhibitor - Everolimus, but there are several critera for this drug to be covered by health insurance. Renal Angiomyolipoma is a tumor of many types and interpretation is time-consuming. Based on the rise of AI in medical imaging, kidneys and tumors can be automatically segmented. We hoped that AI can be used to compare the images of a patient before and after automatically. Registration (alignment). It is a technique to distinguish difference between two images. Several methods of alignment have been used in the research including Advanced Normalization Tools (ANTS), voxelmorph alignment, and visualizing the results of alignment. I designed to automate the entire tumor interpretation process. We hoped that when the doctor inputs a new image file of a patient, it can be automatically segmented, the tumor can be automatically divided into particles, and the total number of tumors, the location and size of each tumor can be calculated, and compared with the previous image. The integrated clinic for tuberous sclerosis in National Taiwan University Hospital has been open 10 years, and has accumulated clinical datas of more than 100 people. This experiment has a total of 116 patients with a total of 26 clinical characteristics. , Bagging Boost, Random Forest) to predict whether the patient has the risk of renal hemorrhage in the future, and use the above data to see whether the patient's genotype can be predicted. The report also takes a month. It is hoped that by means of machine learning, the patient's genotype can be directly inferred from clinical data and images.