結節性硬化症之腎血管肌脂肪瘤的影像判讀及人工智慧應用

Abstract

腎血管肌脂肪瘤(angiomyolipoma)是大多數結節性硬化症患者終身伴隨的腎臟腫瘤，由於發病年齡早，腫瘤會漸進性的成長，等到患者開始感受到腰圍腫脹、腰痛，甚至到無法平躺、腹痛、血尿等情況時，往往腫瘤已經長到一定大小，且侵犯了腎臟實質，造成腎臟不可逆的傷害，因此醫學影像的追蹤對結節硬化症患者來說是非常重要的。 臺大醫院「結節硬化症整合型門診」特別針對結節硬化症患者所成立的門診，幾乎收集了台灣各地區患者數據，由於腎血管肌脂肪瘤對患者的重要性，且具有潛在出血的風險性，所以本研究用回朔性方式，收集個案臨床數據及影像資料，統計出和腎血管肌脂肪瘤分期和腫瘤出血都和患者年齡、基因、aneurysm、TAE有統計上顯著的相關性，回歸係數結果顯示在腫瘤出血事件情況下，年齡26 ~ 50歲是年齡0 ~ 25歲的3.443倍(p: 0.271)，而年齡51 ~ 75歲是年齡0 ~ 25歲的19.526倍(p: 0.014)；同理，在出血事件下對renal angiomyolipoma stage 4 ~ 6是stage 0 ~ 3的24.137倍(p: 0.005)，而對有伴隨intra-lesion aneurysm是沒有aneurysm的2.094倍(p: 0.316)。另外Youden index計算出腫瘤大小4.47公分（sensitivity:92.3%，spicificity:67.1%)為best cut off point，發現stage 3~4以上是腫瘤出血風險關鍵，TSC1基因變異患者集中在腫瘤分期stage 0~3，而TSC2基因變異患者則各個腫瘤分期都有分佈。 由於腎血管肌脂肪瘤組織多變化的特性，相對應有多樣化的醫學影像呈現，再加上臨床上要對腫瘤分期定義以及腫瘤體積的測量等，整個影像的判讀過程很複雜且冗長的，需要專業且有經驗的醫師來評估。 在有限人力資源下，面對大量累積的醫學影像，本研究希望可以運用人工智慧的技術，深度學習（deep learning）醫學影像的影像辨識(image recognition)、物件偵測（object detection)及分割(segmentation)技術，讓電腦可以有效的協助醫師，在判讀的過程有第二診斷。 第一階段訓練104張MRI影像，使用3D-ResNet18神經網路，測試結果腫瘤分期整體accuracy 60%，kappa 0.49，AUC值介於0.67～0.99；aneurysm判斷整體accuracy 85%，precision 80%，sensitivity 66.67%，specificity 92.86%，kappa值0.625，AUC值0.77。第二階段訓練102張標記MRI影像，使用3D-ResNet18+ASPP演算法，測試結果全腎臟分割平均Dice score 0.856 / IOU 75.57%。第三階段訓練65張標記MRI影像，使用2D-samiseResNet40+CC-attention 演算法，測試結果腎血管肌脂肪瘤分割表現，ground truth平均Dice score 0.771 / IOU 64.98% ; inference平均Dice score 0.580 / IOU 44.01%。 本研究整理過去在台灣結節硬化症患者累積的數據，透過收集、整理及統計分析，確實分析出一些和腫瘤分期、腫瘤出血有相關性的結果，而AI醫療影像辨識的應用在台灣已經逐漸發展，也希望這樣的進步可以提供醫生第二意見，協助醫生在影像的判讀，提升的效率，而對於醫療資源不足的地區，病患能夠一樣享有同品質的影像診斷，這對於罕見疾病的患者而言會是非常有利的協助。

The renal angiomyolipoma is a type of benign hamartoma that may occur sporadically or be associated with tuberous sclerosis complex (TSC). They are likely to be early onset, multiple, large, and bilateral, and are prone to grow and be more aggressive. It causes irreversible damage to the kidneys, so the tracking of medical images is very important for patients with tuberous sclerosis. The tuberous sclerosis complex integrated outpatient clinic of National Taiwan University Hospital collected patient data from almost all regions of Taiwan. Because of the importance of renal angiomyolipoma for patients and the potential risk of bleeding, this study used a retrospective method to collect clinical data and imaging data of cases. It is statistically correlated with the age, gene, aneurysm and TAE of patients with renal angiomylipoma stage and tumor bleeding. Regression coefficient results show that in case of tumor bleeding, The odds of age in 25 ~ 50 y is 3.443 times than age 0 ~ 25 y ( p:0.271 ), the odds of age in 51 ~75 y is 19.526 times than age 0 ~ 25 y ( p:0.014 ), and the odds of renal angiomyolipoma stage 4 ~ 6 is 24.137 times as stage 0 ~ 3 ( p:0.005 ). The best cut off point of angiomyolipoma size is 4.47cm (sensitivity: 92.3%；specificity: 67.1%). Above renal angiomyolipoma stage 3 ~ 4 is the key to the risk of tumor hemorrhage. Patients with TSC1 gene mutation are concentrated in stage 0 ~ 3, while patients with TSC2 gene mutation are distributed in each tumor stage. Renal angiomyolipomas are typically composed of smooth muscle, blood vessels, and adipose tissue. The changing characteristics of renal angiomyolipoma, there should be a variety of medical images, in addition to the clinical definition of tumor stage and the measurement of tumor volume, the interpretation process of the entire image is complicated, and requires professional and experienced physicians to evaluate. The renal angiomyolipoma is a type of benign hamartoma that may occur sporadically or be associated with tuberous sclerosis complex (TSC). They are likely to be early onset, multiple, large, and bilateral, and are prone to grow and be more aggressive. It causes irreversible damage to the kidneys, so the tracking of medical images is very important for patients with tuberous sclerosis. The tuberous sclerosis complex integrated outpatient clinic of National Taiwan University Hospital collected patient data from almost all regions of Taiwan. Because of the importance of renal angiomyolipoma for patients and the potential risk of bleeding, this study used a retrospective method to collect clinical data and imaging data of cases. It is statistically correlated with the age, gene, aneurysm and TAE of patients with renal angiomylipoma stage and tumor bleeding. Regression coefficient results show that in case of tumor bleeding, The odds of age in 25 ~ 50 y is 3.443 times than age 0 ~ 25 y ( p:0.271 ), the odds of age in 51 ~75 y is 19.526 times than age 0 ~ 25 y ( p:0.014 ), and the odds of renal angiomyolipoma stage 4 ~ 6 is 24.137 times as stage 0 ~ 3 ( p:0.005 ). The best cut off point of angiomyolipoma size is 4.47cm (sensitivity: 92.3%；specificity: 67.1%). Above renal angiomyolipoma stage 3 ~ 4 is the key to the risk of tumor hemorrhage. Patients with TSC1 gene mutation are concentrated in stage 0 ~ 3, while patients with TSC2 gene mutation are distributed in each tumor stage. Renal angiomyolipomas are typically composed of smooth muscle, blood vessels, and adipose tissue. The changing characteristics of renal angiomyolipoma, there should be a variety of medical images, in addition to the clinical definition of tumor stage and the measurement of tumor volume, the interpretation process of the entire image is complicated, and requires professional and experienced physicians to evaluate.