- 學位論文
使用深度學習對雙能量肺部電腦斷層影像進行淋巴結轉移預測
Nodal Metastasis Prediction on Dual Energy Lung CT Image using Deep Learning
摘要
淋巴結轉移（nodal metastasis, Nmet）是醫生對於癌症診斷時的臨床主要任務，肺癌會依照腫瘤-淋巴結-轉移（Tumor-Node-Metastasis, TNM）報告來進行分期，其中淋巴結轉移與肺癌的生存和復發最為相關。目前來說，在手術前的淋巴轉移預測仍然是對患者管理手術計劃和做出治療決策的挑戰。對於這項挑戰我們提出了新的深度學習的預測方法，是具有腫瘤尺寸相關的模組來預測每個能階（單位為 keV）的影像和具有主要特徵增強模組多能階融合的深度學習模型，用於肺癌原發腫瘤的淋巴結轉移的識別，其中這些深度學習模型我們結合了放射科醫生和計算機科學的相關知識來進行淋巴轉移預測。在這篇文章中，所提出的深度學習模型是使用通過寶石光譜成像（gemstone spectral imaging, GSI）在肺癌原發性腫瘤的雙能量電腦斷層掃描（computer tomography, CT）上具有不同能階而定制設計的。在第一個部分中，使用每個能階的影像訓練出的最佳模型是由40能階影像數據集所訓練，淋巴轉移預測的準確度為86%，Kappa值為72%。第二部分，多個能階影像融合的最佳模型由較低能階影像的組合（40、50、60、70 keV）所訓練，達到93%的準確率和86%的Kappa值。在實驗中，我們有 11 張不同的單色影像從 40至140能階（間隔為10 keV），和三個不同能階融合數據組合於每位患者：較低能階組合，較高能能階組合（110、120、130、140 keV），以及平均能階的組合（40、70、100、140 keV）。當我們使用40能階影像訓練所提出的模型，發現與其他能階在預測淋巴結轉移有顯著差異，並利用交叉驗證（cross-validation）來解釋較低的能階影像在預測原發腫瘤的淋巴結轉移更有效。當我們使用較低能階影像的組合訓練融合模型時，發現與放射科醫生有顯著差異，準確度更接近病理報告。因此，在多能階組合的部分，我們也使用交叉驗證來證明較低能階融合模型更加穩定，更適合用來從原發性腫瘤預測淋巴結轉移。結果表明，較低能階的雙能量電腦斷層影像顯示出更多的腫瘤血管生成與異質性，這些特徵提供我們所設計的深度學習模型用來預測淋巴結轉移。
並列摘要
Lymph node metastasis, also called nodal metastasis (Nmet), is a clinically primary task for physicians. The survival and recurrence of lung cancer are related to the N-staging from Tumor-Node-Metastasis (TNM) reports. Furthermore, preoperative Nmet prediction is still a challenge for the patient in managing the surgical plan and making treatment decisions. We proposed a novel deep prediction method using per energy level image with a size-related damper block (SR-DB) and a multi-energy level fusion model with a principal feature enhancement (PFE) block for Nmet identification from the primary tumor in lung cancer, which incorporating radiologist and computer science knowledge for Nmet prediction. The proposed models are custom-designed by gemstone spectral imaging (GSI) with different energy levels on dual-energy computer tomography (CT) from a primary tumor of lung cancer. In the first part, the best model using per energy level image is trained by the 40 keV dataset achieves an accuracy of 86% and a Kappa value of 72% for Nmet prediction. In the second part, the energy level fusion model is trained by the lower energy level set (40, 50, 60, 70 keV) that achieves an accuracy of 93% and a Kappa value of 86%. In the experiment, we have 11 different monochromatic images (per energy level) from 40~140 keV (the interval is 10 keV) and take three different energy level fusion datasets: a lower energy level set, a higher energy level set (110, 120, 130, 140 keV), and an average energy level set (40, 70, 100, 140 keV) for each patient. When we used the proposed model of a 40 keV image, there has a significant difference in other energy levels (unit of keV). Therefore, we apply 5-fold cross-validation to explain in the first part that the lower keV is more efficient in predicting the Nmet of the primary tumor. When we used the lower energy level set to train the fusion model, there has a significant difference between the fusion model and radiologists, which shows the lower energy level fusion model has the predicted results closer to the pathology report. Hence, we also apply 5-fold cross-validation in the second part to demonstrate that the lower energy level fusion model is more robust and suitable in predicting the Nmet of primary tumors. The results show that the lower energy level shows more tumor angiogenesis or heterogeneity provided the proposed models to predict Nmet from primary tumors.