整合主動輪廓模型之深度卷積網路以辨識遮罩不完全精準之醫學影像

Abstract

肝癌是目前全球最常見的癌症之一，每年都造成大量的死亡人口。醫生在進行診斷與治療之前，通常必須先了解肝臟和肝腫瘤的狀態，然而判讀醫學影像往往耗時又耗力。正好近年分割模型隨著深度神經網路的再崛起迅速發展，應用在醫學影像上，能幫助醫生快速地獲得輔助診斷資訊、減少誤判，成為重點應用領域。然而分割模型多為監督式學習，其學習效果受遮罩的品質影響甚鉅，且人工圈選時須在效率與細節中取捨，因此使用不完全精準的遮罩來訓練模型為本研究之重點。 為了在使用不完全精準的遮罩時仍能有效提升圖像分割的準確度，本研究提出一新穎的模型訓練架構，首先導入擷取局部能量並限制更新區域的主動輪廓模型來修正遮罩輪廓，讓模型有更精準的訓練目標；接著將影像與品質提升後的遮罩輸入以EfficientNet-b7作為編碼器、DeepLabV3Plus作為解碼器的分割模型，並透過增加輸入以及輸出的連續層數，來模擬醫生觀察上下切面的連續資訊量；最後則是混合使用二元交叉熵、Dice損失係數以及RMI (Region Mutual Information) 來確保局部、全域、形狀等多層次資訊均納入考量的損失函數來進行模型學習。 本研究針對台大醫院的肝癌個案進行分析，將目標重疊率從91%改善至96%，並能微量挑出血管與腫瘤等非肝臟部分，針對圖像邊緣的輪廓也更加貼合肝臟本身，有部分的模型圈選結果甚至經醫生驗證較原先手動的目標遮罩品質更好，獲得全方面的成效提升。

Hepatic cancer is one of the leading causes of death from cancer globally. To perform precise treatments, doctors have to evaluate liver function and tumor status carefully. Conventionally, liver function is assessed with the assistance of analyzing the medical images, which is time-consuming and sometimes ineffective. As the techniques of computer vision advance along with the hyper development of deep neural nets, the analytical efficiency and effectiveness for medical image examination are significantly enhanced with the assistance of segmentation models. Recent studies show that the mainstream segmentation models are constructed via supervised learning methods given proper masks on the images, which means the model performance substantially depends on the target mask precision that determined by the efforts in mask formation. To cope with the imprecise masks while maintaining satisfactory model performance, a novel training framework is proposed by employing the local-based active contour model firstly to improve the mask precision. MR (Magnetic Resonance) images with enhanced mask quality are used to train a deep neural net model, initiated with EfficientNet-b7 as the backbone encoder followed by DeepLabV3Plus as the decoder. Five consecutive slices of images are also utilized in model training to consider the informative continuity within the neighbor slice. The model is finally learned by mixing the binary cross-entropy, Dice loss, and region mutual index loss together to keep the information across distribution, region, and shape of the masks. The proposed framework in the thesis is validated by the liver dataset provided by NTUH (National Taiwan University Hospital) partners. The IoU (Intersection over Union) index is improved from 91% to 96%, in comparison to the state-of-the-art segmentation models. Furthermore, some hepatic vessels and tumors can be detected and removed by our model without prior information. It is even verified by medical doctors that the proposed model can highlight the regions that are easily ignored, and works as beneficial assistance in image diagnosis.