Automatic Image Classification for the Urinoculture Screening Ing. Paolo Andreini Ing. Simone Bonechi DIISM University of Siena December 11th, 2015
Urine Culture Standard Protocol Plate Analysis Sample Collection Seeding Incubation
Possible Advantages Better Accuracy Reduced Costs-Time Action Required Just in Case of Error Results Later Available
Main Goals Colony Strain Classification Bacterial Count ROI Extraction Acquisition Device
Algorithm Pipeline Colonies Strain Classification Bacterial Count Plate Detection Background Subtraction Acquisition
Plate Detection Frame Differencing Hough Transform Live Capture Motion Detection Plate Detection Saved Image
Uriselect 4 Chromogenic Medium Non Selective Opaque Note: our samples have been sown manually
Background Extraction Background Model Background Subtraction Meanshift Segmentation CIE-Lab Color Space To Compensate for Local Background Dishomogeneities
Yellow Colonies Effect of the Base Algorithm Effect of the Local Feature Addition Original Image
Classification Stage Chromogenic Substrate - Uriselect 4 ColiFaecalisKes groupS. Agalactiae ProteusPseudomonasS. AureusCandida
Pre-Classification Red Blue Yellow
Multistage Classification Pre Classification Yellow Colonies Classifier Allow to Distinguish Between the Three Main Groups Blue Colonies Classifier Allow to Distinguish between the Three Main Groups Red Colonies can be just recognized
FeatureExtraction Background Subtraction Meanshift Segmentation a,b (CIE-Lab) Original Image To Compensate for Local Dishomogeneities
Pre Classification Architecture
Pre-ClassificationResults Analysed Segments Number13292 Correctly Classified % Errors % (MLP Structure 2 6 3) MLP multilayer perceptron
Blue Colonies' Classification Background Subtraction Pre Classification “Sure” Background Blue colonies + Background GrabCut Algorithm
Blue Colonies' Classification GrabCut Algorithm Effect Background Subtraction Effect Original Image
Blue Colonies Classification Results Analysed Segments Number7122 Correctly Classified % Errors % MLP multilayer perceptron (MLP Structure 2 3 6 3)
Bacterial Count Expressed in UFC/ml (Number of Microorganisms per Milliliter of Urine) The Evaluation Scale is Logarithmic Represents an Estimation of the Infection Severity
Single Colonies Detection Mask Foreground Single Colonies Min Enclosing Circle for each Connected Component th(Circle Area/ Component Area)
Slightly Overlapping Colonies Searching for “seeds” Ellipse SelectionResult Selection Based on a Score Matrix Concavity/Convexity of Contour Estimation
CandidaRecognition Original Petri Plate Edge Detection Colonies Detection Searching for Not Overlapping Colonies Based on Sobel Operator
ChromID CPS Chromogenic Medium Non Selective Semi transparent Note: the samples have been sown automatically
Automatic Seeding BioMérieux PREVI™ Isola Samples are spread circularly Noise Elements on plate
Circular Seeding
Pre Processing Pre-Processing Written text Removal Label Removal ROI estraction
Written Text Removal Color Enhance Color Model Generalized Hough transform Post processing Sobel based Edge detection Template model Selection by: Rotation Position Dimension
Written Text Removal Source ImageText Removed
Written Text Removal Text can be occluded, is it useful to find it in this case?
Written Tet Removal Results Total Number of Image499 Correctly Identify 433 Errors 66 Accuracy 86,78% Accuracy in Infected Plate: 75,45% (160/212) Accuracy in non Infected Plate: 95,12% (273/287)
LabelRemoval Morphological FilteringMin Enclosing Rectangle Otsu Thresholding (find two distribution)
Anonimize the Plate Blur the patient's name for privacy reasons
Label Removal Results Total Number of Image499 Correctly Identify 499 Errors 0 Accuracy 100%
BackgroundRemoval The culture ground appearance is modeled by MOG
Infection Severity Estimation Max ConcentrationImage is Divided in SectorsPre-Processing
Infection Severity Estimation Probe the image counter-clockwise The spread angle gives the estimation
Infection Severity Estimation Results Total Number of Image499 Correctly Identify 477 Errors 22 Accuracy 95,4% Positive – Negative Classification Results PositiveNegative Confusion Matrix
Infection Severity Estimation Results Total Number of Image212 Correctly Identify 176 Errors 36 Accuracy 83,01% Results Confusion Matrix
Infection Classification The infections appearance is modeled using MOG
Infection Classification The image is segmented accordingly
Infection Classification In the uncertain regions the posterior probability is low Those regions can be ignored
Coming Soon Improve the segmentation performance using “local” informations
Coming Soon Adapt to Different Types of Culture Ground and Seeding Techniques