應用通道注意力機制和加權特徵融合之ConvNeXt網路於甲狀腺超音波影像亢進診斷

Abstract

甲狀腺亢進是一個伴隨嚴重症狀，並有甲狀腺癌潛在高風險的疾病。在早期診斷出甲狀腺亢進，可以減緩甲狀腺癌的病情進展並得以及早治療。甲狀腺超音波具有非侵入性與無放射線攝取的特質，因此常被用於甲狀腺亢進的診斷。然而使用超音波影像進行診斷，非常仰賴放射科醫師的經驗與資質，而可能導致誤診的發生。因此可以使用一個具有自動診斷功能的電腦輔助診斷(computer-aided diagnosis, CADx)系統，來協助放射科醫師進行甲狀腺亢進診斷。 近年來，卷積神經網路(Convolutional Neural Network, CNN)因其在疾病診斷上的優異表現，而被廣泛用於開發電腦輔助診斷上。因此本研究提出了一個基於卷積神經網路的電腦輔助診斷系統，應用於甲狀腺超音波影像並進行甲狀腺亢進診斷。本系統包含了影像預處理與甲狀腺亢進分類。在影像預處理中，會自輸入影像中提取甲狀腺體區域，並調整至固定影像大小。接著，經過預處理的影像會作為所提出的分類模型的輸入，並由模型進行甲狀腺亢進的分類。所提出的分類模型取名為ECA-VCR-ConvNeXt，是修改自ConvNeXt架構而來。為了自原ConvNeXt萃取出包含更多資訊的特徵，加入了高效通道注意力(efficient channel attention, ECA)模塊。此外，為了提升ConvNeXt所萃取特徵的品質，我們提出了向量通道校準(vectorial channel recalibration)模塊並用於實作出加權特徵融合(weighted feature fusion)機制。 在實驗中，我們使用了共851張甲狀腺超音波影像，其中包含了419張甲狀腺亢進影像和432張甲狀腺正能影像，來評估我們診斷系統的表現。根據實驗結果顯示，本研究提出的電腦輔助診斷系統可達到86.37%的正確率、84.49%的靈敏性、88.19%的特異性和0.9046的曲線下面積。實驗結果顯示所提出的系統在甲狀腺亢進的診斷上有非常優秀的能力。

Hyperthyroidism is a disease with severe symptoms and a potentially high risk of thyroid cancer. An early hyperthyroidism diagnosis can slow the disease progression of thyroid cancer and obtain early treatment. Thyroid ultrasound with non-invasivity and no radiation load is a common approach for hyperthyroidism diagnosis. However, the high dependence on the experience and qualification of radiologists could lead to misdiagnosis. Therefore, a computer-aided diagnosis (CADx) system with automatic diagnosis ability could assist radiologists. Recently, the convolutional neural network (CNN) has been widely used in developing the CADx system because of its outstanding performance on disease diagnosis. Thus, a CNN-based CADx system was proposed in this study for hyperthyroidism diagnosis on thyroid ultrasound images. The CADx system was composed of image preprocessing and hyperthyroidism classification. In image preprocessing, the thyroid region was extracted from each input image and resized to a fixed resolution. Then, the preprocessed images were fed into the proposed model, ECA-VCR-ConvNeXt, for hyperthyroidism classification. The proposed ECA-VCR-ConvNeXt was modified from a CNN architecture named ConvNeXt. To extract more informative features through the ConvNeXt, the efficient channel attention (ECA) module is embedded. Furthermore, the vectorial channel recalibration (VCR) module was proposed and utilized to introduce the weighted feature fusion to the ConvNeXt for enhanced quality of features. In experiments, a total of 851 thyroid ultrasound images, including 419 hyperthyroidism images and 432 euthyroidism images, were utilized to evaluate the performance of the proposed system. According to the experiment results, the proposed CADx system could achieve 86.37% accuracy, 84.49% sensitivity, 88.19% specificity, and 0.9046 AUC. These results indicated that the proposed CADx system had an excellent capability in hyperthyroidism diagnosis.