利用深度卷積網路於乳房超音波預測乳癌淋巴轉移

Abstract

乳腺癌是女性癌症死亡的第二大原因。儘管癌症治療方法之有所進展，轉移性乳腺癌仍然會導致高死亡率，準確預測和早期檢測乳腺癌轉移狀態對乳腺癌患者的預後是相當重要的。腋窩淋巴節轉移之狀態是評估癌症分期和決定乳腺癌患者治療策略的重要指標。因此，我們開發了一種基於卷積神經網絡之新型電腦輔助預測系統，該系統使用乳房超音波影像來預測腋窩淋巴節之狀況。首先，採用Mask R-CNN模型從整張乳房超音波偵測出腫瘤之位置並切割出腫瘤區域。在獲得腫瘤區域後，我們提取腫瘤周圍區域並使用DenseNet模型進行腋窩淋巴節狀況之預測。在我們的實驗中，共有153例乳腺腫瘤患者, 其中有59例腋窩淋巴節有轉移，94例腋窩淋巴節沒有轉移，用於評估我們提出的方法。根據我們實驗的結果，最佳預測表現是使用包含腫瘤和腫瘤周圍3公厘之區域。其準確性，靈敏度和特異性分別為81.05％（124/153），81.36％（48/59）和80.85％（76/94），ROC曲線下的面積為0.8054。我們所提出的電腦輔助預測系統，結合原發腫瘤的腫瘤和腫瘤周圍區域，在乳癌患者中，可以有效的預測腋窩淋巴節轉移的狀況。

Breast cancer is the second leading cause of cancer death in women. Despite the improvement in cancer therapy, metastatic breast cancer still leads to a high mortality rate. Accurate prediction and early detection of breast cancer metastasis status are important for the prognosis of breast cancer patients. Axillary lymph nodes (ALN) status is an important indicator in assessing cancer staging and deciding the treatment strategy for patients with breast cancer. Consequently, we developed a novel computer-aided prediction (CAP) system based on convolutional neural networks (CNN) using breast ultrasound to distinguish the ALN status. At first, the Mask R-CNN model was used to detect the tumor location and segment the tumor region from the whole US image. After obtained the tumor region, we extracted the peritumoral region and used the DenseNet model to predict the ALN status. In the experiments, 153 patients with breast tumor composed of 59 cases with ALN metastasis and 94 cases without ALN metastasis, used to evaluate our proposed method. According to results, the best prediction performance was using tumor region with the peritumoral region (3mm), the accuracy, sensitivity, specificity, and AUC are 81.05% (124/153), 81.36% (48/59) and 80.85% (76/94), and 0.8054. In summary, the proposed CAP model combines the primary tumor and peritumoral region, which can be an effective method to prediction the ALN status in patients with breast cancer.