Upload
asli-yazagan
View
37
Download
3
Embed Size (px)
Citation preview
Analysis of Inter-Rater Reliability of the Mammography
Assessment after Image Processing
IWBBIO 2015Granada / SPAIN
Aslı YAZAĞANR.T.E.U
Rize / TURKEY
15.04.2015 IWBBIO 2015 Aslı Yazağan
15.04.2015 IWBBIO 2015 Aslı Yazağan
15.04.2015 IWBBIO 2015 Aslı Yazağan
15.04.2015 IWBBIO 2015 Aslı Yazağan
15.04.2015 IWBBIO 2015 Aslı Yazağan
IMAGE PROCESSING
15.04.2015 IWBBIO 2015 Aslı Yazağan
15.04.2015 IWBBIO 2015 Aslı Yazağan
This visual difference hides any valuable diagnostic information for
the detection of abnormalities ?
15.04.2015 IWBBIO 2015 Aslı Yazağan
Introduction
Subjective interpretation of mammograhy images. Mammography Decision Support System (MDSS) Important not losing diagnostic information after image
processing for system performance
15.04.2015 IWBBIO 2015 Aslı Yazağan
Classification of Mass
Mass Detection
Image Enhancement( Median Filtering, Wavelet Based Thresholding, CLAHE, Anizot-opik
Diffusion Filtering )
MDSS System Performance
Introduction
Radiologists’ consistency in their interpretation before and after image processing?
How to understand the concept of radiologists’ disagreement and observer variability.
Two measurements for reliability:1. Intra-rater Reliability
One person should come out with the same results on every repetition of the test.2. Inter-rater Reliability
The degree of agreement among raters
15.04.2015 IWBBIO 2015 Aslı Yazağan
Objective of the Study
To test the consistency of radiologists’ evaluation before and after image enhancement
Research Question
Does image enhancement step cause any diagnostic data loss that affects radiologists’ decision?
15.04.2015 IWBBIO 2015 Aslı Yazağan
Methodology
Collection of Mammography ImagesEvaluation of ImagesObserver Consistency
15.04.2015 IWBBIO 2015 Aslı Yazağan
Methodology > Collection of Mammography Images
Database of the Mammographic Image Analysis Society (MIAS) was used.
50 mammogram from MIAS database 20 images include tumors 20 images include calcification 10 normal images
15.04.2015 IWBBIO 2015 Aslı Yazağan
Methodology > Evaluation of Images
2 female – 3 male radiologists screened and marked selected images before and after image enchancement process.
3 months interval between two sessions to reduce the influence of learning effects on the results.
15.04.2015 IWBBIO 2015 Aslı Yazağan
Methodology > Evaluation of Images
Radiologists decided
Is there a mass or calcification the number of abnormalities the abnormality level of the suspicious region
Evaluation results from all radiologists were combined for each image as two scores as indications of doctors’ decisions before and after image enhancement.
The closeness of the scores indicates the consistency of the radiologists’ decision on that image.
15.04.2015 IWBBIO 2015 Aslı Yazağan
Methodology > Observer Consistency
The consistency of radiologists’ evaluation is assessed by inter-rater reliability statistics.
Cronbach’s Alpha, Sperman Correlation, Pearson's R and Kappa statistics were used to evaluate radiologist reading consistency.
15.04.2015 IWBBIO 2015 Aslı Yazağan
Results
15.04.2015 IWBBIO 2015 Aslı Yazağan
Three radiologists had high consistency especially according to Cronbach's Alpha.
Two radiologists showed a moderate consistency. Inconsistency might depend on other factors such as experience, age, sex, psychological situation etc.
That mass detection performance is more consistent than calcification detection. It is decided to use other image processing methods for calcification images.
Discussions
15.04.2015 IWBBIO 2015 Aslı Yazağan
Diagnostic performance studies depend on radiological problem domain image processing techniques that used in the study.
How to infer a general judgment that image processing improves or reduces radiologists’ diognostic quality.
Wide and long term projects should be planned for generic usage that helps radiologists’ interpretation of images
Image Processed ImageApplying an Image processing technique Improves Radiologists’
diagnostic quality ??
Conclusion
15.04.2015 IWBBIO 2015 Aslı Yazağan
This is a part of MDSS project. Important not losing diagnostic information after image
enhancement for accurate classification of abnormalities for MDSS.
Diagnostic information lost is tested by radiologists. This study focuses on the consistency of radiologists’
evaluation before and after image enhancement.
Similar studies needs to be conducted before designing medical decision support systems.
Methodology > Evaluation of Images
Radiologists decided the number of abnormalities the abnormality level of the suspicious region
Rows represent abnormality counts for each image.
Columns represent the degree of the probability of being malign of those regions
15.04.2015 IWBBIO 2015 Aslı Yazağan