以機器學習協助腸道支架設計之自動化

Abstract

大腸直腸癌高居常見癌症第三名，而約有三成大腸直腸癌患者會併發腸道阻塞，導致腹痛、腹脹甚至腸壞死。傳統上發生腸道阻塞時會進行開腹手術，然而此介入性治療有多項風險與缺點，因此在腸道支架概念被提出後，透過置放支架撐開阻塞區域的治療方法，便成為另一種可行的方案。 腸道阻塞條件因個體差異會有很大的不同，包含狹窄程度、阻塞區域長度等，因此若要套用同一款支架設計於所有腸道阻塞情況並不妥善，開發多種設計以提供不同個體、不同狹窄程度適當的徑向強度是較常見的選擇。傳統的支架設計開發，從繪圖、雛型品製造至測試，需耗費數天才能完成一次設計迭代。透過有限元素分析軟體的協助，工程師得以在電腦中模擬評估支架，進而將設計迭代縮短至數小時。然而由於設計的初次及後續調整十分仰賴經驗與直覺，雖能藉由模擬大幅降低單次迭代所需時間，仍需大量迭代數天方能達到設計目標，而此時間成本在需開發多種設計、甚至客製化設計時會更加明顯。 本研究開發全新的腸道支架設計與參數架構，並以Python語言整合Solidworks製圖、Abaqus模擬、機器學習預測，完成腸道支架之設計自動化。本研究連動製圖軟體與有限元素分析軟體，將有限元素模型建構流程自動化，高效取得模擬結果作為訓練數據，最終利用類神經網路及支持矢量機等機器學習方法建構模型，以預測徑向強度、縮短量、應變等重要的支架性質。在模型訓練完成後，輸入目標即可取得建議的支架設計參數，取代有限元素模擬、經驗調整設計以及大量迭代，將數天縮短為數秒，快速進入驗證及微調階段。

Colorectal cancer ranks as the third most common cancer, and about 30 percent of patients with colorectal cancer will develop intestinal obstruction, causing abdominal pain, abdominal distension and even intestinal necrosis. Traditionally, laparotomy is performed when intestinal obstruction occurs. However, this interventional treatment has multiple risks and disadvantages. Therefore, after the concept of colonic stent was proposed, treatments to open the blocked area through stent placement becomes another feasible solution. The conditions of intestinal obstruction vary greatly between individual patients, including the degree of stenosis and the length of the obstruction area, etc. Therefore, it is not appropriate to apply the same stent design to all intestinal obstruction situations. It is a common choice to develop multiple designs to provide appropriate radial strength for different individual patients and different degree of stenosis. It takes several days to complete a design iteration for the traditional bracket design and development, from drawing and prototype manufacturing to testing. With the help of the finite element analysis software, engineers can simulate and evaluate stents in the computer, thus reducing the design iteration to several hours. However, since the initial and subsequent adjustments of the design rely heavily on experience and intuition, although the time required for a single iteration can be greatly reduced by simulation, the design goal still need days to be achieved after a large number of iterations, and the time cost will be more obvious when multiple designs or even customized designs need to be developed. This study developed a new design and parameter architecture of colonic stent, using Solidworks, Abaqus simulation and machine learning prediction in Python language to complete the design automation of colonic stent. In this study, by connecting cartographic software and finite element analysis software, automating the construction of finite element models, the simulation results were obtained efficiently, which are collected as training data. Then, this study constructed models by machine learning methods such as artificial neural network and support vector machine, so as to predict the important stent properties like radial strength, shorten length and strain. After the completion of model training, the recommended stent design parameters can be obtained by entering the target outputs. The process replaces finite element simulation, experience-based design adjustment and large number of iterations, shorten computation time from several days to a few seconds. The idea helps us enter the validation and fine-tuning stage quickly.