Yonsei-TAIL / KHD2020

Sinusitis classification
7 stars 5 forks source link

Related Works #1

Open ChoiDM opened 3 years ago

ChoiDM commented 3 years ago
  1. Deep Learning in Diagnosis of Maxillary Sinusitis on Conventional Radiograph (@seunghyunni)

  2. Improvement diagnostic accuracy of sinusitis recognition in paranasal sinus X-ray using multiple deep learning models (@Impku)

  3. Deep-learning classification using convolutional neural network for evaluation of maxillary sinusitis on panoramic radiography (@HanBJaaaaaa)

  4. Performance of deep learning object detection technology in the detection and diagnosis of maxillary sinus lesions on panoramic radiographs (@ChoiDM)

Impku commented 3 years ago
  1. Improvement diagnostic accuracy of sinusitis recognition in paranasal sinus X-ray using multiple deep learning models 관련 정리
    • Task: x-ray를 이용한 부비동 축농증 식별
seunghyunni commented 3 years ago
  1. Deep Learning in Diagnosis of Maxillary Sinusitis on Conventional Radiograph 정리
  1. train data: 8000장
  2. validation data: 1000장
  3. test data: - external test set(n = 140), geographic external test set ( n = 200)
  1. 전처리: 다이콤 형식 X-ray 영상들을 7x7 (cm) --> 70mm * 70mm 사이즈로 센터 크롭 이후 left maxillary sinuses 에 해당하는 영상을 horizontal flip하여 추가 데이터로 이용함. 이후, 모든 이미지들을 224x224사이즈로 bilinear interpolation 통해 upsampling하여 사용
  2. augmentation: random scaling, random translation, random contrast modification.
  1. 10개의 residual block과 하나의 fc layer로 이루어진 ResNet 사용
  2. conv layer마다 batch normalization 사용
  3. cross-entropy loss 사용 / RMSProp optimizer 사용
  4. lr 0.01에서 시작하여 1000 step마다 0.94 decay rate으로 줄음
  5. batch 48
  6. Xavier weight initialization

image

HanBJaaaaaa commented 3 years ago
  1. Deep learning classification using convolutional neural network for evaluation of maxillary sinusitis on panoramic radiography

    -

    Task: panoramic X선(파노라마 영상) 영상을 이용하여 (maxilary sinusitis)상악동염을 진단하고 diagnostic performance를 clarify.

    Data: 내부데이터:

    • training data : 400명의 정상인과 염증 환자 데이터를 사용(상악동염 위치에 따라 4그룹(양쪽 다 정상, 좌, 우 감염, 양측 감염)으로 분류하고 남자 50,여자 50으로 나누어짐).

    • test data : 30명의 양쪽 모두 정상인 사람과 과 60명의 unilateral한 염증환자 데이터 사용

    Pre-processing:

  2. crop: ROI를 200x200로 crop

  3. data agumentation: Infan View SW의 필터들을 이용하여 smaple 6000개까지 늘림(로테이션은 사용하지 않음), Table 3 참조

    -

    Model: Alexnet CNN 라이브러리 V5.0 사용

    epochs : 200 사용 환경 :Ubuntu OS v. 16.04.2 with 128 GB of memory and 11 GB of GPU memory (NVIDIA GeForce GTX 1080 Ti).

    -

    Results Accuracy DL : 87.5%, Radiologist : 89.6%, (Dental) Resident : 76.7%

    Sensitivity

    DL : 86.7%, Radiologist : 90.0%, Resident : 78.3%

    Specificity

    DL : 88.3%, Radiologist : 89.2%, Resident : 75.0%

    Positive predictive value

    DL : 88.1%, Radiologist :89.6%, Resident : 75.0%

    Negative predictive value

    DL:86.9%, Radiologist : 89.9%, Resident : 77.6%

    AUC

    DL : 87.5%, Radiologist : 89.6%, Resident : 76.7%

    -

    Conclusion

    DL 시스템이 방사선 전문의에 가까운 퍼포먼스를 보였고, 치과 Resident보다는 훨씬 높은 퍼포먼스를 보임( Dental resident가 전문의인지 전공의인지는 파악이 잘 안되네..)

    -

    특이사항

    • 기본적인 Alexnet(V5.0 라이브러리)를 사용하였고, 딥러닝 모델들간의 비교가 아닌 Human doctor/dentist 와 비교한 논문.

    • augmentation을 rotation하지 않고 brightness, contrast, sharpness, gamma correction, blur, middle value 등의 필터들을 사용하여 진행함.

첨부. Table3

2020년 9월 19일 (토) 오후 12:29, seung hyun Hwang notifications@github.com님이 작성:

  1. Deep Learning in Diagnosis of Maxillary Sinusitis on Conventional Radiograph 정리

    -

    Task: Water's view radiopraph( X- 선 빔이 궤도 선에 대해 45 ° 각도를 이루는 X-ray 영상) 에서의 maxillary sinusitis(염증성 축농증) 자동 진단

    Data: 직접 레이블해서 생성한 데이터

  2. train data: 8000장

  3. validation data: 1000장

  4. test data: - external test set(n = 140), geographic external test set ( n = 200)

    • Pre-processing:
  5. 전처리: 다이콤 형식 X-ray 영상들을 7x7 (cm) --> 70mm * 70mm 사이즈로 센터 크롭 이후 left maxillary sinuses 에 해당하는 영상을 horizontal flip하여 추가 데이터로 이용함. 이후, 모든 이미지들을 224x224사이즈로 bilinear interpolation 통해 upsampling하여 사용

  6. augmentation: random scaling, random translation, random contrast modification.

    • Model:
  7. 10개의 residual block과 하나의 fc layer로 이루어진 ResNet 사용

  8. conv layer마다 batch normalization 사용

  9. cross-entropy loss 사용 / RMSProp optimizer 사용

  10. lr 0.01에서 시작하여 1000 step마다 0.94 decay rate으로 줄음

  11. batch 48

  12. Xavier weight initialization

    • Results:
  13. temporal external test set: AUC 0.93 / Sensitivity 0.77 / Specificity 0.94

  14. geographic external test set: AUC 0.88 / Sensitivity 0.56 / Specificity 0.99

    • Discission: Class Activation Map (CAM)을 뽑아본 결과, 가장 sinusitis 가 잘 발생하는 지역에선 mucus(점액), 또는 mucosal thickening 이 잡혔다. 반대로 non-sinusitis 의 경우에는 bony wall (코뼈) 부분이 잡혔다.

[image: image] https://user-images.githubusercontent.com/57741623/93658122-c0242a00-fa73-11ea-9e47-ad2fa303dea6.png

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/CCIDS/KHD2020/issues/1#issuecomment-695157681, or unsubscribe https://github.com/notifications/unsubscribe-auth/AH6HX3RVKAF6CNQLTESL5ZTSGQQSVANCNFSM4RRZP2WQ .