| 摘要: | 大腸癌是全球最普遍且致死率最高的惡性疾病之一，根據國民健康署癌症登記最新的資料顯示，我國大腸癌的發生人數已連續15年位居世界之首，平均每年都增加約1萬5、6千名新診斷患者。
大腸癌的發展與息肉的形成密切相關，因此透過內視鏡檢查來提早發現並進行息肉切除手術就可以有效預防大腸癌的發生。對於內視鏡醫生來說，是否能夠在檢查期間準確區分結腸息肉類型並做出明智的治療決策至關重要。
圖像增強內視鏡檢查技術，包含色素內視鏡檢查和窄帶成像，用於增強息肉表面的凹坑模式或微血管外觀，這些被強化的視覺特徵，例如：色素分布、血管密度、病變範圍和隆起程度等，能夠大幅提升醫生對結腸息肉類型的診斷能力。然而出於成本考量，窄帶成像內視鏡的普及性在某些臨床環境中可能會受到限制。
為了替內視鏡醫師提供更低的設備成本和有效的評估，並解決由白光息肉病變成像中紋理特徵模糊導致之卷積神經網絡模型預測能力有限的問題，本文提出了一種新的輸入特徵表示方法，該方法結合了紋理特徵、頻譜特徵和高提升濾波器，用於基於白光成像的結腸息肉多類分類。本文還提出了一個新的模型架構CBAMNet，通過結合CBAM與其他多項知名的正歸化技術，進一步提升了白光息肉分類的準確率。
實驗結果表明，本研究所提出的紋理特徵對 CNN 於白光息肉圖像之分類性能有著顯著的影響 (p<0.05)，經由對測試資料進行5-Fold交叉驗證，我們所提出的特徵使得六個最先進的CNN模型的平均F1分數提升了14.0%。透過結合我們的方法和模型，我們成功地將白光下的多類別結腸息肉分類F1分數提高至87.0%。研究結果突出了結合紋理特徵對於提高結腸息肉白光成像分類準確率的重要性以及CBAMNet的分類潛力。;Colorectal cancer is one of the most common and deadliest malignancies worldwide. According to the latest data from the National Health Agency′s Cancer Registry, the incidence of colorectal cancer in Taiwan has been the highest in the world for 15 consecutive years, with an average increase of approximately 15,000~16,000 newly diagnosed patients each year. The development of colorectal cancer is closely related to the formation of polyps. Therefore, early detection and polypectomy through endoscopic examinations can effectively prevent the occurrence of colorectal cancer. For endoscopists, the ability to accurately differentiate colonic polyp types and make informed treatment decisions during examinations is crucial.
Image-enhanced endoscopic techniques, including chromoendoscopy and narrow-band imaging, are used to enhance the surface patterns or microvascular appearance of polyps. These enhanced visual features, such as pigment distribution, vascular density, lesion extent, and protrusion degree, can significantly improve the diagnostic capabilities of endoscopists in identifying colonic polyp types. However, the widespread use of narrow-band imaging endoscopy may be limited in certain clinical settings due to cost considerations.
To provide endoscopists with lower equipment costs and effective evaluations, and to address the limited predictive capabilities of convolutional neural network (CNN) models caused by the blurred texture features in white-light polyp imaging, this study proposes a novel input feature representation method. This method combines texture features, spectral features, and high-boost filters for multi-class classification of colonic polyps based on white-light imaging. Additionally, a new model architecture called CBAMNet is introduced, which further enhances the accuracy of white-light polyp classification by combining CBAM with other well-known normalization techniques.
Experimental results demonstrate that the texture features proposed in this study have a significant impact on the classification performance of CNN in white-light polyp images (p < 0.05). Through 5-fold cross-validation on the test data, our proposed features improve the average F1 score of six state-of-the-art CNN models by 14.0%. By combining our method and model, we successfully increase the multi-class colonic polyp classification F1 score to 87.0% under white-light conditions. The research findings highlight the importance of combining texture features to improve the accuracy of white-light colonic polyp imaging classification, as well as the classification potential of CBAMNet. |
