考慮到台灣女性子宮頸癌的高發病率,增強預防和篩查方法顯得至關重要。由於細胞學檢查與子宮頸切片之間常存在差異,加上手術後傷口復原可能引起的炎症反應會影響抹片的判讀,導致診斷不足的問題變得更加突出。因此,治療後的抹片追蹤變得非常重要。所以本研究旨在針對抹片異常的患者,減少「異常切片的偽陰性」和「不必要的抹片及切片檢查」。過去的研究顯示,傳統抹片的準確率僅為 55-80%(平均約 75%),而近期使用 VGG 和 ResNet 等 AI 模型的準確率已超過 75%,達到與抹片相當的水平,但這些模型較為笨重很難在資源較為匱乏的地區派上用場,所以本研究希望能在邊緣式裝置下達到相同的效果,因此選用了基於 ImageNet 權重的輕量級 MobileNetV2 預訓練模型,並在其後加上 CBAM 模塊,開發出一個針對子宮頸轉化區類型和子宮頸上皮內瘤變的分類辨別模型。本研究使用了 Intel 於 Kaggle舉辦的陰道鏡子宮頸轉化區種類分類比賽中的資料集進行訓練,模型在僅有 1.79M參數量的情況下,實現了 79.52%的準確率和 79.46%的 Macro-F1 分數,並將其作為預訓練模型應用於國立成功大學醫學院附設醫院的子宮頸 CIN 分類資料中。結果顯示,使用此預訓練的方式在有限的資料和資源下的效果可以與其他較為繁重的模型訓練結果相媲美,達到 78.2%的準確率和 79.5%的 Macro-F1分數。研究結論強調了AI 在改進子宮頸癌篩查準確性方面的潛力,特別是在減少偽陰性和不必要的抹片檢測中有顯著貢獻。模型的預測結果令人鼓舞,證明了輕量型模型在準確分類子宮頸病變和幫助識別治療後變化方面的可行性。未來本研究將進一步添加更多的成大醫院病患資料,以鞏固模型在各種臨床環境中的應用性。

 

 

 

Given the high incidence of cervical cancer among women in Taiwan, enhancing prevention and screening methods was crucial. Discrepancies often existed between cytology and cervical biopsy results, and inflammation from post-surgical wound healing could affect smear interpretations, exacerbating the issue of underdiagnosis. Therefore, post-treatment smear follow-up became increasingly important. This study aimed to reduce false negatives in abnormal biopsies and minimize unnecessary smears and biopsies in patients with abnormal smears. Past studies showed that the accuracy of traditional smears ranged from 55% to 80% (averaging around 75%). Recently, AI models like VGG and ResNet achieved accuracy rates exceeding 75%, comparable to smear tests. However, these models were cumbersome and challenging to deploy in resource-limited areas. Thus, this study aimed to achieve similar results on edge devices by selecting a lightweight MobileNetV2 pre-trained model based on ImageNet weights, enhanced with a CBAM module. This model was developed to classify cervical transformation zone types and cervical intraepithelial neoplasia (CIN). The study utilized the dataset from the Intel & MobileODT Cervical Cancer Screening competition on Kaggle for training. The model, with only 1.79 million parameters, achieved 79.52% accuracy and a Macro-F1 score of 79.46%. It was then applied as a pre-trained model to CIN classification data from National Cheng Kung University Hospital, resulting in 78.2% accuracy and a Macro-F1 score of 79.5%. The study concluded that the lightweight model demonstrated comparable performance to heavier models, especially in reducing false negatives and unnecessary smears. The encouraging results underscored the feasibility of using lightweight models for accurately classifying cervical lesions and aiding in post-treatment identification. In the future, this study will incorporate additional patient data from National Cheng Kung University Hospital to further strengthen the model's applicability in various clinical settings.