content based medical image indexing and retrieval using a fuzzy compact composite descriptor

19
CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR Savvas Chatzichristofis and Yiannis Boutalis Department of Electrical & Computer Engineering Democritus University of Thrace – Greece Signal Processing, Pattern Recognition and Applications SPPRA 2009 Presenter: Savvas A. Chatzichristofis

Upload: guesta2cfc

Post on 10-May-2015

1.768 views

Category:

Technology


8 download

TRANSCRIPT

Page 1: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Savvas Chatzichristofis and Yiannis BoutalisDepartment of Electrical & Computer Engineering Democritus University of Thrace – Greece

Signal Processing, Pattern Recognition and Applications SPPRA 2009

Presenter: Savvas A. Chatzichristofis

Page 2: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

• Compact Composite Descriptors (CCD) are global image descriptors capturing more than one features at the same time, in a very compact representation.

Natural ImagesCEDDFCTH

Artificial ImagesSpCL

Medical ImagesBTDH

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 3: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Proposed Descriptor BTDH

•This descriptor uses brightness and texture features in one compact vector.

•Its size does not exceed 48 bytes per image.

•This characteristic makes the descriptor appropriate for use in large medical image databases.

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 4: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Proposed Descriptor BTDH

•To extract the brightness information, a fuzzy unit classifies the brightness value of the image’s pixels into 8 clusters.

•The texture information embodied in the proposed descriptor is a Fuzzy approach of the Tamura Directionality Histogram.

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 5: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Descriptor Implementation

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 6: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Pre-Filtering unit• Auto brightness correction

This method is partially inspired by the HVS (Human Vision System). It particularly adopts some of the shunting characteristics of the on-center off-surround networks, in order to define the response function for a new artificial center-surround network.

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 7: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Pre-Filtering unit• Edge enhancement

A coordinate logic filter (CLF) ‘OR’ is applied to the image. This filter enhances the edges of the image. Thus, it aims to help the texture information extraction unit to reach weaker texture alterations.

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 8: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Brightness information extraction unit • This unit purposes to

classify the brightness of the pixels into 8 clusters using a fuzzy classification system.

• The fuzzy system output is an 8 bin histogram.

• The centre of these clusters was calculated using Gustafson-Kessel algorithm on a sample of 1000 (8 bit greyscale) medical images.

Brightness Classification System

Fuzzy system output

V(A)

V(B)

C(0) C(1) C(2) C(3) C(4) C(5) C(6) C(7)3.18 22.68 54.00 90.13 125.80 162.57 202.25 243.64

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 9: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Texture Information extraction unit • For every image block

entered into the texture information extractor unit, an 16-bin histogram that describes the directionality of the image block is extracted.

• Directionality histogram is a graph of local edge probabilities against their directional angle.

• The fuzzy system output is an 16 bin histogram. Fuzzy system output

V(B)

V(A)

Texture Classification System

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 10: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Descriptor Structure• The descriptor’s structure has n=16 regions determined

by the Directionality Unit. Each Directionality Unit region contains m=8 individual regions defined by the Brightness Unit. Overall, the proposed descriptor contains n X m = 128 bins.

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 11: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Descriptor Implementation

1. Auto brightness correction 2. Edge enhancement 3. Divided into 3X3 Blocks 4. Directionality Form n=2 5. Brightness Form m=3 6. Bin(19) is Activated

The procedure is repeated for all the Blocks. On the completion of the process, the descriptor's histograms bin values are normalized within the interval [0,1]

To restrict the proposed descriptor storage requirements, the bin values of the descriptor are quantized for binary representation using a three bits/bin quantization.

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 12: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Quantization• For each image entered into the system, the proposed

descriptor is extracted. This descriptor is separated into the 16 texture regions. The value of each bin of the descriptor is assigned to one of the values [0,7] according to the minimum distance of the value from one of the eight entries in the corresponding row of the quantization table.

000 001 010 011 100 101 110 1113.6E-04 2.3E-03 4.5E-03 7.2E-03 1.1E-02 1.6E-02 3.1E-02 5.7E-013.2E-04 1.9E-03 3.7E-03 5.7E-03 8.4E-03 1.3E-02 2.4E-02 5.3E-013.3E-04 1.9E-03 3.5E-03 5.3E-03 7.6E-03 1.1E-02 1.6E-02 4.0E-023.7E-04 2.1E-03 4.1E-03 6.3E-03 8.8E-03 1.2E-02 2.0E-02 6.9E-023.2E-04 1.8E-03 3.4E-03 5.4E-03 7.9E-03 1.1E-02 1.7E-02 3.9E-023.1E-04 1.7E-03 3.3E-03 5.3E-03 7.8E-03 1.1E-02 1.8E-02 5.6E-023.6E-04 2.0E-03 3.9E-03 6.1E-03 8.6E-03 1.3E-02 2.1E-02 7.8E-023.6E-04 2.1E-03 4.0E-03 6.2E-03 9.0E-03 1.3E-02 2.4E-02 1.9E-013.7E-04 2.3E-03 4.4E-03 7.0E-03 1.0E-02 1.6E-02 2.9E-02 2.4E-014.2E-04 2.5E-03 4.9E-03 7.6E-03 1.1E-02 1.9E-02 5.3E-02 6.8E-013.6E-04 2.1E-03 3.9E-03 6.3E-03 9.3E-03 1.4E-02 2.7E-02 4.4E-013.7E-04 2.2E-03 4.3E-03 6.9E-03 1.1E-02 1.6E-02 3.5E-02 3.8E-013.3E-04 2.0E-03 3.8E-03 6.2E-03 9.4E-03 1.4E-02 2.8E-02 3.7E-012.7E-04 1.8E-03 3.6E-03 5.7E-03 8.9E-03 1.3E-02 2.6E-02 3.6E-013.0E-04 2.0E-03 3.9E-03 6.1E-03 8.8E-03 1.3E-02 2.6E-02 3.7E-013.2E-04 2.1E-03 4.2E-03 6.8E-03 1.0E-02 1.5E-02 3.1E-02 4.3E-01

Bin(19)=0.006

n=2

Bin(19)=3The final size of the proposed descriptor is

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 13: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Similarity Measure• The similarity between the images was calculated using

the non-binary Tanimoto Coefficient

• Where xt is the transpose vector of X.• In the absolute congruence of the vectors the Tanimoto coefficient is

equal to 1, while in the maximum deviation the coefficient tends to zero.

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 14: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Experiments• The proposed descriptor has been implemented and is

available as open source library under GNU - General public License (GPL) in the image retrieval system img(Rummager) and the on line application img(Anaktisi).

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 15: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Experiments• To evaluate the performance of the proposed descriptor, the objective

measure called ANMRR is used.• The experiments were carried out in a group of 5000 images with 120

queries. • A set of ground truth images that are most relevant to the query were

identified. The ground truth data is a set of visually similar images.

Descriptor ANMRR

Proposed Method Using Tanimoto 0.272

Proposed Method Using Jensen-Shannon 0.283

Proposed Method Using Euclidian 0.287

Tamura Directionality Histogram 0.321

Gabor Vector 0.328

MPEG7: Edge Histogram 0.381

Gray Value Histogram 0.448

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 16: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Experiments on IRMA 2005 Medical Image Database• The IRMA database consists of

10000 annotated radiographs taken randomly from medical routine at the RWTH Aachen University Hospital-Germany. The images are separated into 9000 training images and 1000 test images. The images are subdivided into 57 classes.

Descriptor MAPProposed Descriptor 28.1Gabor vector 27.7Gray value histogram 26.1Gabor histogram 25.2inv. feature histogram (mon.) 24.4inv. feature histogram (relational) 24.1LF patches signature 23.0Tamura Directionality Histo. 21.6LF SIFT global search 20.9LF patches global 17.6global texture feature 16.4LF SIFT signature 10.9MPEG7: edge histogram 10.9

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 17: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Demonstration

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Page 18: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Conclusions• The experimental results showed that the proposed

descriptor can be used for the retrieval of medical images more successfully than the Tamura Directionality Histogram.

• The proposed method can be used as part of a broader retrieval system that uses more characteristics, replacing the Tamura Directionality Histogram.

Page 19: CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR

Download the img(Rummager) application from http://www.img-rummager.com

Thank YouΕυχαριστώ Πολύ

CONTENT BASED MEDICAL IMAGE INDEXING AND RETRIEVAL USING A FUZZY COMPACT COMPOSITE DESCRIPTOR