本研究主要針對陰道鏡的子宮頸影像進行子宮頸癌偵測和分類。我們與其他相關研究不同的地方在於我們發現除了一般常用的判讀特徵之外,病灶邊緣白化的型態是一個可能的特徵。所以我們在分類之前先針對病灶的邊緣的型態進行分群的動作,讓每一個分群的病灶其邊緣有著一致的型態,然後針對每一個分群抽取白化和邊緣灰階對比的程度作為特徵來訓練分類器,進行子宮頸癌的分類。我們使用434張影像來進行實驗,測試分群前後的分類正確率,我們測試了其中一個分群證實分群之後的正確率達到97.17%,此研究之成果可作為醫師診斷癌化等級時的輔助參考,更準確的診斷出癌化等級,使病人接受最適當的醫治,減少因誤判所造成之遺憾。
This study is focus on the detection and classification of the cervical cancer by the colposcopic cervical images. We found that the pattern of the edge changes could also be a feature for the interpreter of lesion upon cervix. In this study, we first cluster lesions according to the pattern of the edge changes. Then we train classifier for each cluster with the features of whiteness and gradient of contour. We collected 434 images for the experiments and test one of the clusters for the precision and compared to the results generated from all images. In the experimental results, we demonstrated that after clustering the precision will be 97.17%. Our research can help doctor to improve their diagnosis and the quality of treatment as well.