Presentazione standard di PowerPoint

1
Automatic Image Classification for the
Urinoculture Screening
Ing. Paolo Andreini Ing. Simone Bonechi DIISM ?
University of Siena
December 11th, 2015
2
Urine Culture Standard Protocol
Sample Collection
Seeding
Incubation
Plate Analysis
3
Possible Advantages
Better Accuracy
Action Required Just in Case of Error
Results Later Available
Reduced Costs-Time
4
Main Goals
ROI Extraction
Colony Strain Classification
Acquisition Device
Bacterial Count
5
Algorithm Pipeline
Acquisition
Plate Detection
Background Subtraction
Colonies Strain Classification
Bacterial Count
6
Plate Detection

Live Capture
Motion Detection
Plate Detection
Saved Image
Frame Differencing
Hough Transform
7
Uriselect 4
Opaque
Non Selective
Chromogenic Medium
Note our samples have been sown manually
8
Background Extraction
Background Subtraction
Background Model
Meanshift Segmentation
CIE-Lab Color Space
To Compensate for Local Background
Dishomogeneities
9
Yellow Colonies
Effect of the Base Algorithm
Effect of the Local Feature Addition
Original Image
10
Classification Stage
Coli
Faecalis
Kes group
S. Agalactiae
Proteus
Pseudomonas
S. Aureus
Candida
Chromogenic Substrate - Uriselect 4
11
Pre-Classification
Red
Blue
Yellow
12
Multistage Classification
Red Colonies can be just recognized
Pre Classification
Blue Colonies Classifier
Allow to Distinguish Between the Three Main Groups
Allow to Distinguish between the Three Main Groups
Yellow Colonies Classifier
13
Feature Extraction
Original Image
Background Subtraction
Meanshift Segmentation
a,b (CIE-Lab)
To Compensate for Local Dishomogeneities
14
Pre Classification Architecture
15
Pre-Classification Results
MLP multilayer perceptron
(MLP Structure 2?6?3)
16
Blue Colonies' Classification
Background Subtraction
Sure Background

GrabCut Algorithm

Pre Classification
Blue colonies Background

17
Blue Colonies' Classification
GrabCut Algorithm Effect
Background Subtraction Effect
Original Image
18
Blue Colonies Classification Results
MLP multilayer perceptron
(MLP Structure 2?3?6?3)
19
Bacterial Count
Represents an Estimation of the Infection Severity
Expressed in UFC/ml (Number of Microorganisms
per Milliliter of Urine)
The Evaluation Scale is Logarithmic
20
Single Colonies Detection
Foreground
Single Colonies
Mask
Min Enclosing Circle for each Connected Component
th(Circle Area/ Component Area)
21
Slightly Overlapping Colonies
Searching for seeds
Ellipse Selection
Result
Selection Based on a Score Matrix
Concavity/Convexity of Contour Estimation
22
Candida Recognition
Original Petri Plate
Edge Detection
Colonies Detection
Based on Sobel Operator
Searching for Not Overlapping Colonies
23
ChromID CPS
Semi transparent
Non Selective
Chromogenic Medium
Note the samples have been sown automatically
24
Automatic Seeding
BioMérieux PREVI Isola
Samples are spread circularly
Noise Elements on plate
25
Circular Seeding
26
Pre Processing
Written text Removal
Pre-Processing
Label Removal
ROI estraction
27
Written Text Removal
Selection by Rotation Position Dimension
Color Model
Template model
Color Enhance
Generalized Hough transform
Post processing
Sobel based Edge detection

28
Written Text Removal
Source Image
Text Removed
29
Written Text Removal
Text can be occluded, is it useful to find it in
this case?
30
Written Tet Removal Results
Accuracy in Infected Plate 75,45
(160/212) Accuracy in non Infected Plate 95,12
(273/287)
31
Label Removal
Morphological Filtering
Min Enclosing Rectangle
Otsu Thresholding (find two distribution)
32
Anonimize the Plate
Blur the patient's name for privacy reasons
33
Label Removal Results
34
Background Removal
The culture ground appearance is modeled by MOG
35
Infection Severity Estimation
Max Concentration
Image is Divided in Sectors
Pre-Processing
36
Infection Severity Estimation
Probe the image counter-clockwise
The spread angle gives the estimation
37
Infection Severity Estimation Results
Positive Negative Classification
Results
Confusion Matrix
38
Infection Severity Estimation Results
Results
Confusion Matrix
39
Infection Classification
The infections appearance is modeled using MOG
40
Infection Classification
The image is segmented accordingly
41
Infection Classification
In the uncertain regions the posterior
probability is low
Those regions can be ignored
42
Coming Soon

Improve the segmentation performance using
local informations
43
Coming Soon

Adapt to Different Types of Culture Ground and
Seeding Techniques