基於深度學習與蒙地卡羅方法之小鼠肺癌劑量分析系統之研究

Deep Learning and Monte Carlo Method Based Dose Analysis System for Lung Cancer in Mice

蔡豐任(Feng-Ren Tsai)

指導教授 ： 吳順吉

國立清華大學/原子科學院/工程與系統科學系/碩士(2021年)

摘要

近年來隨著生醫影像大量的數位化，人們可開發出不同的演算法對它們進行分析以解決對應的問題。本研究旨在建立一套自動化的硼中子捕獲治療(Boron Neutron Capture Therapy，BNCT)劑量分析系統，先藉著卷積神經網路(Convolutional Neural Networks，CNNs)自動且快速地進行電腦斷層掃描(Computed Tomography，CT)影像之腫瘤偵測，接著使用蒙地卡羅方法(Monte Carlo Method)進行劑量分析，用以評估病患接受BNCT之治療前最佳的照射時間，避免正常組織接收到過多的劑量。
系統的建構分成兩階段，第一階段對原始CT影像進行前處理，並訓練輸入串接(Input Cascaded)多尺度分析模型對小細胞肺癌小鼠CT影像進行腫瘤辨識，建立一個高準確率的辨識系統，第二階段藉由腫瘤辨識模型所預測出的腫瘤區域建置均質化小鼠肺癌體素模型，並使用蒙地卡羅計算程式(Monte Carlo N-particle transport code，MCNP)計算經由清華水池式反應器(Tsing Hua Open Pool Reactor，THOR)照射，小鼠體內器官組織在不同T/N Ratio下之劑量分佈。在腫瘤辨識的部分，經過訓練的腫瘤辨識模型對腫瘤區域有超過90%的成功率，預測整筆3D影像(512張2D切片)只需約1小時的時間，能夠正確且迅速地找出影像中之腫瘤組織。在劑量分析的部分，我們對MCNP的輸出檔進行分析，發現在T/N Ratio大於2時，可以在不超過正常組織的耐受劑量的情況下殺死腫瘤細胞，且治療可在1小時內完成。

關鍵字

硼中子捕獲治療 ； CT影像 ； 肺癌 ； 卷積神經網路 ； 蒙地卡羅 ； 劑量分析

並列摘要

With the increasing digitization of a large number of biomedical images, people can develop different algorithms to analyze them for different purposes. This study aims to establish a treatment planning system for Boron Neutron Capture Therapy (BNCT). The system first uses the proposed convolutional neural networks (CNNs) to automatically detect tumors from computed tomography (CT) images, followed by Monte Carlo Method for dose analysis for best irradiation time to prevent normal tissues from receiving too much dose.
The construction of the system is divided into two stages. In the first stage, the original CT images are pre-processed, and the Input Cascaded multi-scale analysis model is trained to perform tumor delineation on CT images of mice with small cell lung cancer to establish a high-accuracy delineation system. The second stage is to build a homogenized mouse lung cancer voxel model based on the tumor areas delineated by the delineation system and use Monte Carlo N-particle transport code (MCNP) to calculate the dose distribution of organs and tissues in mice under different T/N Ratio through Tsing Hua Open Pool Reactor (THOR) irradiation. For tumor delineation, the obtained model achieved a success rate of more than 90%. Moreover, It took approximately 1 hour to process the entire image stack (512 2D slices). As for dose analysis, we analyzed the output file of MCNP and found that when T/N Ratio was more than 2, tumor cells could be effectively dealt with without exceeding the tolerated dose of normal tissues.

並列關鍵字

BNCT ； CT images ； lung cancer ； convolutional neural network ； Monte Carlo ； dose analysis

參考文獻

[1] R. F. Barth, A. H. Soloway, and R. G. Fairchild, "Boron neutron capture therapy for cancer," Scientific American, vol. 263, no. 4, pp. 100-107, 1990.

[2] R. F. Barth, A. H. Soloway, R. G. Fairchild, and R. M. Brugger, "Boron neutron capture therapy for cancer. Realities and prospects," Cancer, vol. 70, no. 12, pp. 2995-3007, 1992.

[3] R. F. Barth, J. A. Coderre, M. G. H. Vicente, and T. E. Blue, "Boron neutron capture therapy of cancer: current status and future prospects," Clinical Cancer Research, vol. 11, no. 11, pp. 3987-4002, 2005.

[4] H. R. Withers, J. M. Taylor, and B. Maciejewski, "Treatment volume and tissue tolerance," International Journal of Radiation Oncology* Biology* Physics, vol. 14, no. 4, pp. 751-759, 1988.

[5] T. Aihara and N. Morita, "BNCT for advanced or recurrent head and neck cancer," in Neutron Capture Therapy: Springer, 2012, pp. 417-424.