基於Mask R-CNN角膜反射亮點偵測之自動眼底攝影系統

做眼底攝影檢查，可篩檢出視網膜病變、青光眼初期症狀或其他慢性病。但通常只有專業眼科診所或醫院眼科門診有專業醫生或是眼底攝影儀器可以做眼底檢查，然而不是每個人都會定期至專業診所或醫院檢查，導致眼底檢查率低。因此於本論文中，提出自動眼底攝影系統，將所有眼底攝影流程自動化，不需專業人員操作，使用者便可完成眼底攝影，希望透過此系統的開發來降低眼底攝影檢查的門檻，以增加其便利性及普遍性。自動眼底攝影系統中透過瞳孔定位模組及自動控制模組自動化眼底攝影拍攝流程，並設計簡易的自動眼底控制器，方便使用者操作。更採用深度學習Mask R-CNN網路來學習不同大小、角度與位置的眼睛角膜反射亮點，以增加瞳孔定位模組的準確性。實驗結果顯示Mask R-CNN角膜反射亮點，經訓練與測試後，其辨識平均精準率(mAP)為91.2%、準確率(Accuracy)為93.73%、敏感度(Sensitivity)為95.43%以及特異度(Specificity)為88.81%。此外，自動眼底攝影系統實際測試結果，平均操作時間則為66.7秒，相對平均一位病患就可省下護理人員或視光師30秒左右的時間；問卷調查的使用者的總滿意度平均為4.6。

Fundus photography examination can screen for retinopathy, initial glaucoma symptoms and other chronic diseases. However, only professional eye clinics or hospital ophthalmology clinics have professional doctors and equipment to exam fundus. Furthermore, everyone will not go to professional clinics or hospitals regularly, and it results a low fundus examination rate. To solve these problems, an automatic fundus photography system is proposed to automate all fundus photography processes in this paper. Users can complete fundus photography without professional operation. It is hoped that the development of this system will lower the threshold for fundus photography examination and increase its convenience and universality. The pupil positioning module and the automatic control module are used to automate the process of capturing fundus photograph, and a simple automatic fundus controller is designed for user convenience. The architecture is based on a Mask R-CNN network, which is used to learn corneal reflection points of different sizes, angles and positions to increase the accuracy of the pupil positioning module. The experimental results show the mean average precision (mAP) of recognition is 91.2%, accuracy is 93.73%, sensitivity is 95.43% and specificity is 88.81% in the Mask R-CNN corneal reflection points. In additional, the average operation time is 66.7 seconds. In general, the system proposed in this paper saves a lot of time (approximately thirty seconds) for the caregiver or optometrist. And he total user satisfaction of the questionnaire survey about automatic fundus photography system is about 4.6.