使用深度學習自動選取超音波影像及診斷

Abstract

乳癌為十大癌症之一，也是女性癌症致死率第二高的主因，若及早對其進行偵測、診斷及治療，就可以將病程延緩甚至治癒。在臨床篩查(Clinical screening)上，超音波檢測有助於評估乳房腫瘤的良惡性，其中手持型超音波(Handheld ultrasound, HHUS)是目前評估乳癌的主流工具。醫師在檢測有疑慮的區域時會鉅細靡遺地記錄掃描過程中可能有異常的部分，並且針對高度可疑的腫瘤做活體組織檢查(Biopsy)，然而詳細地篩查會使得成像數量繁多，因而造成後續醫師在檢視時有不小的負擔。因此，我們提出一個自動挑片暨診斷系統以減輕醫護人員的重擔。本系統分為二階段：在第一階段中，我們使用基於ImageNet的預訓練Transformer模型，自每位患者的影像序列中挑出較有疑慮的影像；而在第二階段則使用改良過的卷積神經網路(Convolutional neural network, CNN)模型對這些有疑慮的影像做良惡性的診斷，該模型以預訓練的ResNeSt-50為基底，結合了全域二階池化模組(Global Second-order Pooling, GSoP)，為了更進一步地獲得更泛化的結果，我們也於該模型中引入了對比中心損失函數(Contrastive center loss)。 在本研究中，我們使用了807位病人來評估我們所提出的系統，醫生對這些病人拍攝了至少5張以上的影像序列(平均影像數為44張)，並且特別挑選出有疑慮的影像。在第一階段實驗，依照模型給予的評分，取用前一及前五高分的影像能涵蓋醫生所標註有疑慮的影像的正確率分別為74.35% (600/807)和97.27% (785/807)；而在第二階段則使用每位病人前五高分的所有影像來評估其良惡性，總計有4035張影像，這些影像涵蓋兩種類別：分別是良性影像3,421張(包含僅有乳房正常組織而沒有腫瘤的影像)及具有惡性腫瘤之影像614張。實驗結果表明，提出的方法的診斷準確度、靈敏度、特異性和ROC曲線下面積(AUC)分別為79.85%、80.13%、79.80%和0.8641。兩個階段的實驗結果顯示了提出的自動挑片暨診斷系統可以減少人員挑片的負擔，並在臨床上提供診斷的第二意見。

Breast cancer is one of the top ten cancers and the second most common cause of death in women worldwide. The disease progression can be postponed or even cured if detected, diagnosed, and treated in the early stage. In general, ultrasound is beneficial in clinical screening to assess the benignity and malignancy of breast cancers, and the use of handheld ultrasound (HHUS) remains the mainstream tool for breast cancer assessment today. When investigating a suspicious region in the breast, operators meticulously record the abnormalities in the scanning process and suggest a biopsy examination for a highly suspicious lesion. However, a detailed screening records many slices, which might increase the review burden for clinicians. Hence, we proposed a computer-aided automatic slice selection and diagnosis system to relieve the burden on the medical staff. It consisted of two stages: In the first stage, we used a Transformer-based model to select suspicious slices from each patient's slice sequence. In the second stage, we employed a modified convolutional neural network (CNN) model to assess whether suspicious slices were benign or malignant. It contained a modified pre-trained ResNeSt-50 model embedded with Global Second-order Pooling (GSoP) blocks. Besides, we also introduced the contrastive center loss to increase the model's generalizability. In this study, we used a total of 807 patients to evaluate our proposed methods, and experienced clinicians recorded the screening slice sequence (range: 5 to 117 slices, average: 44 slices) and indicated which were of interest. In the first stage, the selection accuracy based on top-1 and top-5 scores was 74.35% and 97.27%, respectively. In the second stage, we utilized the top-5 slices from 807 patients to assess their benignity and malignancy, in which there were 3,421 benign slices (normal breast tissue and no tumor) and 614 slices containing malignant tumors, summing to 4,035 slices. The experimental results demonstrated that the accuracy, sensitivity, specificity, and AUC of diagnosis were 79.85%, 80.13%, 79.80%, and 0.8641, respectively. The two experiment results showed that the proposed system could reduce personal strain and provide a second diagnostic opinion for the clinic settings.