Progress Seminar 2018. 6.26 선석규

Neural-rim Segmentation 연구 진행 상황 보고 녹내장 진단 (안과 Pf. 박기호) 입술 병변 분류 (피부과 Pf. 정진호) 재난 대응 (응급 의학과 Pf. 신상도) MER 분석 (신경외과 Pf. 백선하) 녹내장 예후예측 OCT Classification 시야검사지 판별 Neural-rim Segmentation 2주전 계획 Segmentation 알고리즘 구현 반대쪽 Recording 신호 이용하여 분석 연구 결과 Method작성하여 전달 Image feature와 OCT 장비상의 추가 Feature 이용 데이터 변경 후 재 진행 및 안과전달 U-Net 설계 및 적용 British Journal of Dermatology Revision 도착 PAPR착용부 개선 및 추가 장비 제작 주제 마무리 문제점 및 대책 실패 accuracy 떨어짐 목표 및 계획 이미지만 이용 필드테스트 준비

Reviewer1 Page 10, Ln 53: It is "344 image training set". The authors have specified previously this set as a test set. Please clarify. Similar statements appear also in the Figure captions. Please check carefully. 논문 내용 : The performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image training set were 0.827  의미 전달이 잘못 되어서 수정함 SNUH의 testset 이 344장이나, train image가 344인 것 처럼 이 부분에만 잘못 표현됨

Reviewer1 2. In addition to Fig.3, examples where the DL-NN did not perform well should be included.  Grad-CAM 관련 문제 : 잘 안된 것도 첨부 예정

Reviewer1 3. During the observer study the authors collected only binary responses from the physicians (malignant, benign). Therefore, the authors were able to plot only single operating points on the ROC plot (with and without aid).  However, a disadvantage of this approach is that it is very difficult to demonstrate that an observer operates on a different improved AUC curve when evaluating with aid. The observer with aid may simply operate on the AUC curve without aid but at a different operating point with changed specificity and sensitivity for example. Therefore no improvement will be observed when evaluating with aid. Better design would be to ask the observers for the likelihood of lesion being malignant. In this way a AUC curves for the observers with and without aid will be possible to be generated. This need to be discussed as a limitation in the discussion section.

Reviewer1 B에서 ROC Curve와 DCNN의 도움을 받기 전 후 차이를 겹쳐서 그리는 것의 비교가 문제가 있음 도움을 받기 전과 후가 sensitivity와 specificity의 변화가 있으나 좋아진 건지는 잘 모르겠다는 의견 Likelihood를 전후에 계산해서 ROC Curve를 다시 그리라고 제안함 (전 후의 ROC Curve를 그리고 모델과 비교)

Reviewer2 직접적인 수정사항은 아님. It should be noted that the number of malignant lesions in the non-SNUH datasets was small and essentially restricted to SCC and Bowen’s disease. Nonetheless, the demonstration of generalizability of the results to this outside source of images is important.  Extra validation을 좀 더 했으면 좋겠다는 내용.. I would caution the authors against suggesting in their discussion that such algorithms could/should be applied to individuals in social media settings outside of the medical ‘contract’- as this raises considerable privacy and consent issues. Social media에 올리는 표현이 개인정보 issue가 될 수 있음

Reviewer2 While AI interpretability is currently a hot topic and very relevant to the clinical application of AI, it is not clear what the Grad-CAM images add to this report. As I understand it, they were not provided as data in the reader study and there is no way to assess their utility based on the small number of images provided. I'd suggest omitting them. 어떤 이미지에 Grad-CAM을 적용한 것인지 기술 하고,(SNUH의 Test data임)  적은 양의 이미지에서 유용성에 대한 평가 방법이 없기 때문에, Grad-CAM을 빼는 것을 추천.