Exploiting eInfrastructures for Medical Image Storage and ...dsie10/presentations... · MAMMOGRAPHY...
Transcript of Exploiting eInfrastructures for Medical Image Storage and ...dsie10/presentations... · MAMMOGRAPHY...
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
Exploiting eInfrastructures for Medical
Image Storage and Analysis: A Grid
Application for Mammography CAD
Raul Ramos / CETA-CIEMAT
Miguel Angel Guevara / INEGI-FEUP
1
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
OUR WORK
INTRO TO CADMAMMOGRAPHY
Outline
1. Mammography
2. CAD systems and techniques
3. Validation and relevance
4. Current Status and Issues
5. Motivation and goals
6. Framework description
7. Results
8. On going and future work
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
1 Mammography
Mammography is of major impact of woman
health
• Improve survival rates, reduce health costs
• Improving accuracy (TP, FP, FN, TN)
Digital mammography
•SFM: Screen Field Mammography
•FDM: Full Digital Mammography
•Different image properties.
•Each better for different scenarios
IT Systems: PACS + DICOM
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
1 Mammography
Target lessions (BIRADS atlas)
masses calcifications
architectural distorsions assimetries
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
1 Mammography
BIRADS 2003 Assesment Categories
• Category 0: Needs additional evaluation and/or mammograms
• Category 1: Negative
• Category 2: Benign Finding
• Category 3: Probably Benign Finding – Initial short interval
follow-up suggested
• Category 4: Suspisious Abnormality – Biopsy should be
considered
• Category 5: Highly Suggestive of Malignancy – Appropriate
action should be taken
• Category 6: Known Biopsy – Proven Malignancy
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
2 CAD Systems and ML Techniques
CAD on the literature (with some ambiguity)
• CADe Computer Aided Detection: Finds locations
• CADx Computer Aided Diagnosis: Characterizes a designated region
Second Opinion to ASSIST specialists
Approaches used in literature
• Pixel based
• Feature classification
Base technologies
• ML Classifiers
• Image Processing
• Statistical, etc.
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
2 CAD Systems and technologies
Feature classification
ROI SELECTION
SEGMENTATIONFEATURE
EXTRACTIONCLASSIFICATION
IMAGE ENHANCEMENT
img processing
algorithms
(pixel/bit based)
dynamic contours,
semiautomatic
principal
component
analysis
ANNs, GAs, SVMsGUI based,
image recognition
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
3 Validation and result comparison
Datasets are used to train
and/or validate systems
cross-validation, test/train
split for training and
generalization measure
Sensitivity vs specificity
ROC/FROC Curves
Metrics:
• TEST %, Az, TP/FP RATE
TEST% @ FP/IMAGE
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
4 Current status and issues
Many systems in literature
• Real work started in late 80’s
• Feature classification: Mostly by FFBP, SOM and SVMs
• Microcalcifications >90% TP
• Other lesions 60-90%
Datasets are VERY HARD to build
• Medical institutions VERY sensitive to data privacy and availability
• Only two publicly available data sets:
- Mammographic Image Analysis Society (MIAS) DB. UK. 300 cases
- Univ South Florida Digital Database for Screening Mammography. 3000 cases
• Small data sets for training classifiers or validating data
• Sensible to doctor’s mood and population specificities
R2 Image Checker
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
4 Current status and issues
Clinical acceptance
• A few works on medical workflow integration performace improvement
• Integration in hospital IT systems: PACs, Workstations, etc.
• CAD increases detection rate equivalently to second reading (3%-7%)
BASELINE OF TECHNOLOGIES AND METHODOLOGY ESTABLISHED
NEED TO INCREASE THE SIGNIFICANCE OF RESULTS
NEED TO INCREASE ACCEPTANCE AMONG MEDICINE PROFESSIONALS (TRUST AND CULTURE)
eINFRASTRUCTURES ARE NOT USED AT ALL
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
5 Goals and motivation
Our hypothesis:
• Grid features (federation and security) provide an adecquate
framwork to overcome the described issues
• We can effectively exploit computer power and federated
storage capacity of Grid infrastructures to:
1. Build better data sets (large federated storage)
2. Build better classifiers (large federate computing power)
Our goals:
1. Build a pilot Grid based system achieving results
comparable to literature (this work)
2. Use Grid federations to build large data sets
3. Use Grid federations to explore the search space of possible
classifiers (exploit computer power)
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
2 Grid federations
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
6 Our Framework:Digital Repositories Infrastructure
X.509 security authentication
VO Based authorization
Large digitalized content goes to
the Grid federation
Patient data in local DB
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
6 Our Framework: scenario
Deploy the MIAS DB on a pilot DRI Grid
infrastructure
Have doctors marking and diagnosing images
(create datasets) using Grid based tools
Build feature classifiers upon created datasets
• FFBP ANN
• SOM ANN
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
7 Results
100 cases from MIAS segmented by doctors with DRI
Classifier performance
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
8 Conclusion and future work
We have
• reached a level of accuracy comparable to literature.
• validated the usage of Grid infrastructures.
• built an scalable Grid framework, suited to doctors workflows.
We are now ready to exploit the Grid for CAD
Future (and present)
• build a large db of annotated diagnosed mammograms at
Hospital San Joao (600 cases to date)
• massively explore a large diversity of classifiers
- Used about 10 CPU-months to train about 1000 classifiers on a 50
core cluster, used FFBP, RB, GA, SA to train ANNs
• we can systematically reproduce literature results
• define good exploration strategies
G
Raul Ramos – Exploiting Grid Infrastructures for Medical Image Analysis
DSIE ‘10 DOCTORAL SYMPOSIUM ON INFORMATICS ENGINEERING. Porto, Jan 2010
Exploiting Grid Infrastructures for Medical Image Analysis
Exploiting eInfrastructures for Medical Image Storage and Analysis: A Grid Application for Mammography CAD
THANKS!!!
QUESTIONS @ ANY TIME