Title: Automatic Identification of Bacterial Types using Statistical Image Modeling
1Automatic Identification of Bacterial Types
using Statistical Image Modeling
-
-
- Sigal Trattner, Dr. Hayit Greenspan,
- Prof. Shimon Abboud
-
- Department of Biomedical Engineering
- Faculty of Engineering,
- Tel-Aviv University
- In collaboration with Dr. Gabi Teper
- Spring Diagnostics Ltd.
2Introduction
Bacteria are identified as the prime cause for
disease outbreak. The identification of
bacterial types enables us to find a suitable
cure for the disease and to control it.
Bacteriophage-(phage) typing is a method used to
identify bacterial types by determining the
bacteria reactivity to a set of selected phages
currently it is performed manually.
Positive phage reaction
Negative phage reaction
3 Objective of our study
- Identify profiles of bacterial types
- automatically, using computer
- vision and statistical modeling techniques.
- Spot Finding
- Spot Categorization
To date Manual, subjective diagnosis of the
results. Technology is enabling the increase in
the number of phages used for typing (via new
technology of Spring Diagnostics).
4Non-uniformity of background across the image
5Algorithm Methodology
Spot Finding
Preprocessing
Segmentation
Grid
Feature Extraction and Modeling
Preprocessing
Segmentation
Grid
Preprocessing
Segmentation
Grid
Spot Categorization
Phage Profiling
Set of images (a group)
Phage Profile of bacterial type
6 Statistical Modeling Tool(GMM EM)
EM
- Feature Space GMM
- The distribution of a random variable is a
mixture of K Gaussians if its density function
is -
- Parameter set
7- Given a set of feature vectors, the maximum
likelihood estimation of the parameter set - The EM algorithm iterative method to obtain
- increasing the likelihood in each iteration
8Segmentation
Image pixels (Intensity samples per image)
Foreground Background
GMM
EM
Pixel Intensity
- Intensity samples of each image are modeled as a
Gaussian mixture distribution. - Segmentation Probabilistic affiliation of
each pixel to background and foreground
(signal).
9Segmentation
- Each pixel is now affiliated with the most
probable Gaussian cluster. - Probability of x to be labeled as
foreground or background (1 or 0) -
-
- Pixel labeling
10Spot Finding
Preprocessed image
Original image
Signal
Background
GMM generated for the pixel intensity
distribution per image
11Spot Finding
Segmented image
De-rotated image and grid alignment
12Spot Categorization
GMM
Image Spots (dataall images in group )
Feature Vectors (NA, SI)
Probabilistic Categorization to / -
EM
Features
Aarea of signal T area threshold PPerimeter
of signal
13Spot Categorization
- GMM has clearly separated two main modules,
using the chosen features.
-
SI
NE
14Spot Categorization
Probability
Category
15Spot Categorization Results
Correlation between supervised spot
categorization and automatic spot categorization
Auto
M
Auto
-
M
16From Spots to Bacterial Type Profiling
Average Probability of phage G to /-
Phage profiling
j
Spots related to the same phage G
17 Bacterial Type Profiling
Phage Profile
Phage no. 19
18Differentiation across profiles of different
groups
Group 3
Group 2
Group 1
Group 4
Data4 image groups of approx 40 scanned images
and 144 phages checked on bacteria of
Staphylococcus Aureus
19Profile Uniformity
Similarity of phage profiles that are extracted
from a single group
Group 3 260 images and 144 phages checked
20Results Test Set II
Group 1 328 images Group 2 72 images
Uniformity 90
Algorithm
Image batch of group 1
Uniformity 93
Algorithm
Image batch of group 2
Image batch of group 1
Uniformity 82
Algorithm
Image batch of group 2
21Conclusions
- An automated tool is presented for translating
large sets of microbiological visual
information into a probabilistic phage
profile of a bacterial type. - The tool is consistent, objective and robust.
- Statistical based decisions are used.
- The tool may be applied in different domains
such as phage therapy, and the domain of
cDNA Microarray analysis.
22Thank You
23- Expectation step estimate the Gaussian clusters
to which the points in feature space belong - Maximization step maximum likelihood parameter
estimates using this data