- 學位論文
基於深度學習架構Mask R-CNN實現自動化之指甲乾癬檢測系統
Building Automatic Nail Psoriasis Nail Severity Index System Based on Mask R-CNN
摘要
在臺灣，因為健保制度的便利及政府補助相當的看診費用，造成臨床上每位醫師每天有不下百位的門診病患。基於上述原因，在門診時段內每位病患所分配到的看診時間非常有限，若遇到病患需要做詳細的診斷或是醫師較長時間的說明則會使得整體的門診時間超過原本的時段，醫師需要犧牲自己的休息時段將病人看完。所以為了節省診斷的時間，我們將皮膚科臨床上用來診斷指甲乾癬的評分標準-指甲乾癬嚴重程度指標(簡稱NAPSI)利用深度學習的架構之一的Mask R-CNN 實現。透過輸入透過我們設計的標準化拍照設備所拍攝的指甲影像並輸出其經由NAPSI計算後所獲得的分數及指甲乾癬臨床表現的框選、標記。由於目前臨床上對於指甲影像的拍攝並未有類似於皮膚科中常用來診斷皮膚狀況的皮膚鏡的設備，所以本研究針對拍攝指甲的條件以及醫師所需調控的變因在前端設計了一個標準化的黑盒子拍照設備。透過可變的鏡頭縮放、光源強度以及光源角度來獲取最佳的指甲影像，並將這些影像由受過訓練的皮膚科醫療人員針對指甲乾癬的不同特徵作分類後，將其用以作為深度學習架構的訓練集、測試集，最後透過監督式的學習以及Mask R-CNN的架構實現自動化的指甲乾癬嚴重程度指標。
並列摘要
Clinically, dermatologists usually apply Nail Psoriasis Severity Index (abbreviate as NAPSI) to diagnose and assess the severity of nail psoriasis. However, due to subsidy and convenience of health insurance in Taiwan, doctors in outpatient clinics might need to face nearly hundreds of patients on workdays, they are lack in time to diagnose each patient thoroughly. Therefore, we developed a simple, fast and automatic NAPSI evaluation system by utilizing one of the deep learning architecture, mask R-CNN. Besides, a standard photographic system was constructed for data acquisition. It not only assisted doctors in capturing patients’ clinical signs, but also facilitated the collection of normalized datasets. Hence, pre-processing of image data could be omitted before training our model. Expectantly, the system could accelerate the diagnosis process, which implies dermatologists could save time to obtain more disease-related details individually and thus improving their diagnostic accuracy as well as making precise decisions of treatment.