=Paper=
{{Paper
|id=Vol-1178/CLEF2012wn-ImageCLEF-StathopoulosEt2012
|storemode=property
|title=IPL at CLEF 2012 Medical Retrieval Task
|pdfUrl=https://ceur-ws.org/Vol-1178/CLEF2012wn-ImageCLEF-StathopoulosEt2012.pdf
|volume=Vol-1178
|dblpUrl=https://dblp.org/rec/conf/clef/StathopoulosSK12
}}
==IPL at CLEF 2012 Medical Retrieval Task ==
https://ceur-ws.org/Vol-1178/CLEF2012wn-ImageCLEF-StathopoulosEt2012.pdf
IPL at CLEF 2012 Medical Retrieval Task
Spyridon Stathopoulos, Nikolaos Sakiotis, Theodore Kalamboukis
Department of Informatics
Athens University of Economics and Business, Athens, Greece
spstathop@aueb.gr tzk@aueb.gr
Abstract. This article presents an experimental evaluation on using La-
tent Semantic Analysis (LSA) for searching very large image databases.
It also describes IPL’s participation to the image CLEF ad-hoc tex-
tual and visual retrieval for the medical task in 2012. We report on our
approaches and methods and present the results of our extensive experi-
ments applying data fusion on the results obtained from LSA on several
low-level visual features.
Key words: LSA, LSI, CBIR
1 Introduction
The continuous advances of internet and digital technologies, as well as the
rapidly increasing multimedia content used by modern information systems, have
imposed a need for an efficient system for organizing and retrieving content from
large multimedia collections. However, the performance in image retrieval is still
very far from being effective for several reasons: computational cost, scalability
and performance.
In our runs this year we have experimented with Latent Semantic Analysis
(LSA), a technique that, although has been used successfully in many applica-
tions in the domain of text retrieval, [1] it has not experienced similar success
in CBIR. The main reason being the density of the f eatures × images ma-
trix, C, generated in image retrieval, in contrast to textual retrieval where the
term × document matrix is sparse. As a result, complexity cost of SVD is raised
to prohibitively high levels for both, space and computational time.
In this article we give an overview of the application of our methods to ad-hoc
medical retrieval and present the results of our submitted runs. Our efforts this
year were concentrated on applying the LSA method to a number of low-level
visual features and then using data fusion techniques on the SVD transformed
low rank approximation of images to enhance retrieval. We explore a what we
call SVD-bypass technique to factor the feature matrix by solving a much smaller
in size eigenproblem of the term correlation matrix CC T instead of solving the
SVD of matrix C. This method proved to be a much more efficient and scalable
solution for large data sets.
In the next section, we describe our approach and in the following sections
we present the submitted IPL’s runs on textual and visual retrieval with their
2
corresponding descriptions and results. Finally, in the last section we conclude
the remarks of this work with propositions for future research.
2 Visual Retrieval
According to the traditional use of LSA in information retrieval a term-by-
document matrix, C, is first constructed and an SVD analysis is then performed
on this matrix. However as stated before, the feature matrix in the case of image
retrieval, is a dense matrix. This increases the computational costs of the SVD
analysis to prohibitive levels for large image databases. A typical example for
our database this year, for the color layout feature will produce a matrix of size
11288 × 305000 ≈ 30GB (in double precision) which makes the SVD impossible
to solve with our computer resources. In our LSA implementation we solve the
eigenproblem of the feature correlation matrix CC T instead. This matrix, for
a suitable representation of the images is of a moderate size, demanding less
storage and the eigenvalue problem of CC T can be solved much faster than the
SVD factorization of the matrix C. We then approximate the feature matrix
taking only the k-largest eigenvalues and corresponding eigenvectors of matrix
C, for a suitable value of k.
2.1 Preprocessing of the data
It is well known that the representation of a digital image depends on several fac-
tors, from its resolution to color models etc. In a collection of images, it is highly
possible that there will be important variations considering these characteristics.
Thus, each image undergoes through several transformations before the feature
extraction step. In our case we have applied the following transformations.
1. Size normalization. All images are re-scaled to the same size.
2. Transformation to gray-scale images.
3. Tile splitting. Each image is split into equal-sized, non-overlapping cells we
reference to as tiles.
2.2 Feature Extraction and Selection
The vector representation of the images was based on three low-level features of
MPEG-7, Scalable Color (SC) with 64 coefficients per tile, Color Layout (CL)
having 192 coefficients per tile and the Edge Histogram (EH) feature. Experi-
ments on CLEF 2011 image collection showed that, the extraction of the edge
histogram per tile had a negative impact on retrieval performance, thus this fea-
ture was extracted from the whole image instead. All the features were extracted
using the Java library Caliph&Emir of the Lire CBIR system [2].
Finally, a simple histogram with 32 levels of gray colors was extracted from
each tile. To increase the discriminating power of the histogram, we remove the
levels with high frequency and normalize the remaining histogram values for all
3
images. At the same time, all histogram levels with a total frequency above 80%
are considered stop-words and thus, they are removed. We refer to this feature
as Color Selection Histogram (CSH).
2.3 Construction of the feature-correlation matrix CC T
As we have already mentioned the matrix C is full in the case of CBIR and so
it is the matrix CC T . This matrix multiplication is the most intensive part of
the computations and memory demanding. In our implementation we overcome
all these problems by splitting the matrix C into a number of blocks, such that
each block can be accommodated into the memory (C = (C1 , C2 , ..., Cp )) and
calculate CC T by:
X p
T
CC = Ci CiT (1)
i=1
After solving the eigenproblem of the feature-correlation matrix CC T , the
k largest eigenvectors, say Uk , and the corresponding eigenvalues, are selected.
The original feature vectors are then projected into the k − th dimensional space
using the transformation
yk = UkT y (2)
on the original vector representation of an image y.
2.4 Data Fusion
For the data fusion task we used a weighted linear combination of the results
obtained by using the LSA method on different features, as defined by :
X
SCORE(Q, Image) = wi scorei (Q, Image) (3)
i
were scorei denotes the similarity score of an Image with respect to a feature i.
The weight of each feature type is determined as a function of its performance [3].
The wi ’s were estimated by the square of the Mean Average Precision (MAP)
values attained by the corresponding feature on the CLEF ’11 collection [4].
These values are listed in Table 1.
Table 1. MAP Values of each feature for the CLEF ’11 collection.
Feature MAP
Scalable Color 0.0043
Color Layout 0.0133
Color Selection Histogram 0.0023
Edge Histogram 0.0111
4
3 Textual Retrieval
This year’s collection contains a subset of PubMed of 305000 images from PubMed.
A detailed description of the collection in given in the overview paper in [5]. Since
records have the same structure as in previous CLEF collections over the past
years, we followed the same steps in the textual retrieval task as in CLEF 2011.
Each record is identified by a unique figureID, which is associated with: the title
of the corresponding article, the article URL, the caption of the figure, the pmid
and the figure URL. From the pmid we downloaded the MeSH terms assigned
to each article.
Our retrieval system was based on the Lucene 1 search engine. For index-
ing we removed stop-words and applied Porter’s stemmer. For the multi-field
retrieval the weights of the fields were assigned at indexing time. We kept the
same structure of the database as in CLEF 2009, 2010 and 2011. This year we
used only the default scoring function 2 which was best performing at the 2011
CLEF Ad-Hoc retrieval.
Also we use the same weights for the fields as in the last three years [6, 7].
We used two sets of weights: one, that was estimated empirically on the CLEF
2009 collection and a second set where the weights were estimated by the value
of the Mean Average precision estimated on the CLEF 2010 collection.
1
http://lucene.apache.org/
2
http://lucene.apache.org/core/old versioned docs/versions/3 5 0/api/core/org/
apache/lucene/search/Similarity.html
5
4 Experimental Results
4.1 Results from Textual Retrieval
This year we’ve submitted a total of six runs using different combinations of fields
and corresponding weights. In Table 2 we give the definitions of our textual runs
and in Table 3 their corresponding results.
Table 2. Definitions of IPL’s runs on textual retrieval.
Run ID Description
IPL ATCM Abstract, Title, Caption and Mesh terms all in one field.
IPL TCM Title, Caption and Mesh terms all in one field.
IPL A10T10C60M2 Abstract, Title, Caption and Mesh in 4 fields with
weights 1, 1, 6, 0.2 respectively
IPL A1T113C335M1 Abstract, Title, Caption and Mesh in 4 fields with
weights 0.1, 0.113, 0.335, 0.1 respectively
IPL T10C60M2 Title, Caption and Mesh in 3 fields with
weights 1, 6, 0.2 respectively
IPL T113C335M1 Title, Caption and Mesh in 3 fields with
weights 0.113, 0.335, 0.1 respectively
Table 3. IPL’s performance results from textual retrieval.
Run ID MAP GM-MAP bpref p10 p30
IPL A1T113C335M1 0.2001 0.0752 0.1944 0.2955 0.2091
IPL A10T10C60M2 0.1999 0.0714 0.1954 0.3136 0.2076
IPL T10C60M2 0.188 0.0694 0.1957 0.3364 0.2076
IPL TCM 0.1853 0.0755 0.1832 0.3091 0.2152
IPL T113C335M1 0.1836 0.0706 0.1868 0.3318 0.2061
IPL ATCM 0.1616 0.0615 0.1576 0.2773 0.1742
6
4.2 Results from Visual Retrieval
For the visual retrieval task, we’ve also submitted a total of six runs, using differ-
ent values for the parameter k which defines the selected number of eigen-values
and vectors to use for indexing and retrieval. For all runs a data fusion on vari-
ous features was applied, using the weights acquired from runs of each individual
feature with the CLEF’s 2011 collection. In Table 4 we give the definitions of
our runs and in Table 5 their corresponding results.
Table 4. Definitions of IPL’s runs on visual retrieval.
Run ID Description
R1: IPL AUEB DataFusion EH LSA Edge Histogram and LSA with k=20 on 64 tiles
SC CL CSH 64seg 20k for Scalable Color, Color layout Color Selection Histogram
R2: IPL AUEB DataFusion EH LSA Edge Histogram and LSA with k=50 on 64 tiles
SC CL CSH 64seg 50k for Scalable Color, Color layout Color Selection Histogram
R3: IPL AUEB DataFusion EH LSA Edge Histogram and LSA with k=100 on 64 tiles
SC CL CSH 64seg 100k for Scalable Color, Color layout Color Selection Histogram
R4: IPL AUEB DataFusion LSA LSA with k=20 on 64 tiles
SC CL CSH 64seg 20k for Scalable Color, Color layout Color Selection Histogram
R5: IPL AUEB DataFusion LSA LSA with k=50 on 64 tiles
SC CL CSH 64seg 50k for Scalable Color, Color layout Color Selection Histogram
R6: IPL AUEB DataFusion LSA LSA with k=100 on 64 tiles
SC CL CSH 64seg 100k for Scalable Color, Color layout Color Selection Histogram
Table 5. IPL’s performance results from visual retrieval.
Run ID MAP GM-MAP bpref p10 p30
R4 0.0021 0.0001 0.0049 0.0273 0.0242
R3 0.0018 0.0001 0.0053 0.0364 0.0258
R1 0.0017 0.0001 0.0053 0.0227 0.0273
R6 0.0017 0.0002 0.0046 0.0364 0.0212
R2 0.0011 0.0001 0.004 0.0136 0.0136
R5 0.0011 0.0001 0.0039 0.0091 0.0121
7
5 Conclusions and Further work
We have presented a new approach to LSA for CBIR replacing the SVD analysis
of the feature the matrix C (m × n) by the solution of the eigenproblem for the
matrix CC T (m × m). The method overcomes the high cost of SVD in terms
of memory and computing time. In addition, in all the experiments, of which
only a small part was submitted officially this year, the optimal value of the
approximation parameter was less than 50 which makes the method attractive
for fusion with several low level features. Certainly our approach is promising
and has created new research directions that need further investigation. The
image representation has an impact on LSA performance and a more systematic
research on that direction is currently under progress. Also the eigenvalues of the
matrix CC T follow a Zipfian distribution with the k-th largest values been well
separated giving small residual vectors to machine accuracy, which give us an
evidence on the stability of the calculated eigenvectors. More work is currently
underway in order to determine the stability of the proposed method.
References
1. Deerwester, S.C., Dumais, S.T., Landauer, T.K., Furnas, G.W., Harshman, R.A.:
Indexing by latent semantic analysis. JASIS 41(6) (1990) 391–407
2. Lux, M., Chatzichristofis, S.A.: Lire: lucene image retrieval: an extensible java cbir
library. In El-Saddik, A., Vuong, S., Griwodz, C., Bimbo, A.D., Candan, K.S.,
Jaimes, A., eds.: ACM Multimedia, ACM (2008) 1085–1088
3. Wu, S., Bi, Y., Zeng, X., Han, L.: Assigning appropriate weights for the linear
combination data fusion method in information retrieval. Inf. Process. Manage. 45
(July 2009) 413–426
4. Kalpathy-Cramer, J., Müller, H., Bedrick, S., Eggel, I., de Herrera, A.G.S., Tsikrika,
T.: Overview of the clef 2011 medical image classification and retrieval tasks. In:
CLEF (Notebook Papers/Labs/Workshop). (2011)
5. Müller, H., de Herrera, A.G.S., Kalpathy-Cramer, J., Fushman, D.D., Antani, S.,
Eggel, I.: Overview of the imageclef 2012 medical image retrieval and classification
tasks. In: CLEF 2012 working notes, Rome, Ital,2012 (2012)
6. Gkoufas, Y., Morou, A., Kalamboukis, T.: Ipl at imageclef 2011. In: CLEF (Note-
book Papers/LABs/Workshops). (2011)
7. Gkoufas, Y., Morou, A., Kalamboukis, T.: Combining textual and visual information
for image retrieval in the medical domain. The Open Medical Informatics Journal
5 (2011) 50–57