Notice
Recent Posts
Recent Comments
Link
일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | ||
6 | 7 | 8 | 9 | 10 | 11 | 12 |
13 | 14 | 15 | 16 | 17 | 18 | 19 |
20 | 21 | 22 | 23 | 24 | 25 | 26 |
27 | 28 | 29 | 30 |
Tags
- python
- Growth hacking
- 스태킹 앙상블
- 분석 패널
- 인프런
- 그로스 마케팅
- 마케팅 보다는 취준 강연 같다(?)
- WITH CUBE
- splitlines
- 캐글 산탄데르 고객 만족 예측
- pmdarima
- 그로스 해킹
- DENSE_RANK()
- 컨브넷
- 캐글 신용카드 사기 검출
- lightgbm
- 부트 스트래핑
- ImageDateGenerator
- 로그 변환
- 데이터 증식
- ARIMA
- XGBoost
- 데이터 정합성
- tableau
- 그룹 연산
- 리프 중심 트리 분할
- WITH ROLLUP
- 데이터 핸들링
- 3기가 마지막이라니..!
- sql
Archives
- Today
- Total
LITTLE BY LITTLE
[4] 캐글 RSNA 2022 Cervical Spine Fracture Detection - EDA 본문
- Evaluation metric
- 전체 8개의 row
- 경추가 있으면 2, 없으면 1 부여
- 골절이 있으면 14, 없으면 7 부여
- 전체 8개의 row
train shape: (2019, 9)
train bbox shape: (7217, 6)
test shape: (3, 3)
ss shape: (3, 2)
- Tabular data
- Fracture Analysis 골절 분포에 대한 분석
- class is balanced - there is ~ 50%-50% ratio between patients with and without fractures
- C7(19%) - this is the bone with the most frequent injuries
- C3(4%) - this is the bone with the least number of injuries
- 몇몇 환자들은 두 개 이상 골절을 갖고 있음
- 한 환자에 여러 골절이 일어난 경우, 주로 가까이 위치한 경추가 골절된 경우가 많았음 (히트맵으로 상관계수 확인)
- Study ID
- 마지막 부분 (ex. 1.2.826.0.1.3680043.6200 => 6200) 만 unique하므로 편의를 위해서 바꾸면 좋을듯
- 이미지 분석
- DICOM
- 앞의 row와 column은 이미지 size
- Pixel Spacing : pixel size
- Image Position Patient : x,y,z 좌표를 알려준다. 왼쪽 상단에 위치
- Instance Number = Slice 넘버
- 이미지를 보면 어디가 경추인지 모르기 때문에, 후에 segmentation으로 경추만 추출하기
- Segmentation
- 주어진 segmentation은 nii포맷
- Align orientation (512, 512, 331) => (331, 512, 512)
- 경추만 추출된 모습을 볼 수 있다.
- np.unique() 보면 어디 경추인지 알 수 있음
- 4.3%로 UNET이용, 나머지 추출하기
- dcm 이미지를 saggital view로 바꾸기
- DICOM
- meta data Cleaning (필요 컬럼만 추출, 나머지 drop)
- 이미지 사이즈는 다 동일
- content date 컬럼 drop
- 포함시켜야할 컬럼
- Slice Thickness
- Number of slices per scan
- Image Position Patient
- Image Orientation Patient : 방향 정렬 여부 알려줌
- 질문
- Bounding Box 어디 쓸지
- UNet 쓰기 위해서 axial -> saggital view로 바꾸기
- 기타
- for the vast majority of patients with 2 fractures, the 2 fractured cervical vertebrae are next to each other => this information might be useful for postprocessing
- Fracture Analysis 골절 분포에 대한 분석
RNSA - 2D model [Validate] [PyTorch]
Explore and run machine learning code with Kaggle Notebooks | Using data from multiple data sources
www.kaggle.com
'4-2 > 이미지 딥러닝' 카테고리의 다른 글
Comments