以 U 型 Transformer 模型及分組注意力網路作為電腦輔助 EGFR 突變診斷系統於肺部電腦斷層影像

Abstract

肺腺癌是一種常見的組織學亞型肺癌，而它的成因又跟表皮生長因子受器(epidermal growth factor receptor, EGFR)的突變有關。透過早期的診斷跟標靶治療，可以有效地提升EGFR突變腺癌患者的疾病無惡化存活期(progression-free survival, PFS)。因此，儘早去確認EGFR的突變情形對腺癌患者非常關鍵。過去的研究中提出了電腦斷層掃描(computer tomography, CT)特徵以及影像組學(radiomic)特徵來對EGFR突變進行非侵入的預測。然而，這兩種方法存在著缺點。電腦斷層掃描特徵需要人力去定義以及量測每個特徵，而影像組學是基於固定的公式來提取特徵，提取完還需經過特徵選取(feature selection)來篩選出重要的特徵。近年來, 深度學習(deep learning)在醫學影像的任務上發揮了很大的影響，可以在不用人為介入下自動提取多樣且獨特的特徵。因此，本研究提出一個基於卷積神經網路(convolution neural networks, CNN)模型和Transformer模型的電腦輔助診斷系統(computer-aided diagnosis)來提供快速且非侵入式的診斷。 本研究提出的系統包含了影像前處理，肺結節切割和EGFR分類。在影像前處理中，肺結節和周圍組織會從電腦斷層影像中被提取出來，再進行影像大小的調整以及正規化處理。接著，結合UNETR模型和雙頭(dual head)架構的3-D DHeadUNETR切割模型會從處理過的影像中提取肺結節的遮罩。最後，處理後的影像以及對應的結節遮罩會合併匯入到3-D RGA-SANet分類模型來進行EGFR突變狀況預測。這個3-D RGA-SANet分類模型使用分組注意力(split-attention, SA)區塊來整合多分支架構的特徵，以及全局關係感知注意力(relation-aware global attention, RGA)來捕捉長距離的相依關係。實驗結果指出此系統在肺結節分割上可以達到0.8265和0.7105在Dice係數和IoU上，在最後EGFR的分類上可以達到75.00%的正確率、75.70%的靈敏性、74.31%的特異性和0.7731的ROC曲線下面積，證實了本系統可以有效的幫助EGFR的診斷。

Lung adenocarcinoma is a common histological subtype of lung cancer related to epidermal growth factor receptor (EGFR) mutation. Through early diagnosis and target therapy, the progression-free survival (PFS) rate could be increased effectively in EGFR-positive adenocarcinoma patients. Thus, it is essential to identify EGFR mutation status early for adenocarcinoma patients. Non-invasive methods like chest computer tomography (CT) features and radiomic features have been used in previous studies to predict the EGFR mutation status. However, there are drawbacks to these two methods. Chest CT features need physician efforts to define and measure, and radiomic features are based on fixed functions to extract, which are not flexible and need feature selection. Recently, convolution neural network (CNN) has shown a great impact on medical imaging tasks. It could automatically capture various and distinct features without human intervention. Therefore, a computer diagnosis system (CADx) based on CNN and Transformer models was proposed to provide a fast and non-invasive diagnosis for EGFR status. The proposed CADx system in this study included image preprocessing, nodule segmentation, and EGFR classification. First, nodules and surrounding tissue were extracted to volumes of interest (VOIs), and then the VOIs were resized, and normalized in the image preprocessing. Next, the proposed segmentation model, 3-D DHeadUNETR, which incorporated the UNETR model and the dual head structure, was utilized to acquire nodule masks from preprocessed VOIs. Lastly, the preprocessed VOIs and the corresponding nodule masks were concatenated and fed into the proposed classification model 3-D RGA-SANet for EGFR prediction. The proposed 3-D RGA-SANet was constructed with the split-attention (SA) blocks to integrate the feature maps from the multi-branch structure, and relation-aware global attention (RGA) blocks to seize long-distance relationships in spatial and channel information. In the experiments, our proposed CADx system could achieve the Dice coefficient and Intersection over Union (IoU) of 0.8265 and 0.7105 in nodule segmentation and reach the accuracy, sensitivity, specificity, and the area under the receiver operating characteristic (ROC) curve (AUC) of 75.00%, 75.70%, 74.31% and 0.7731 in EGFR classification. The results indicated that our proposed system could assist radiologists in EGFR status diagnosis.