Automatic classification of breast density according

Jordi Freixenet, Arnau Oliver, Joan MartíUniversity of Girona

{jordif, aoliver, joanm}@eia.udg.es

In this work we propose a new algorithm to classify breasts according their internal tissue. Moreover, and in contrast with most of the previous works, the classification follows the BIRADS standard. Our proposal consists on grouping those pixels with similar tissue by using a clustering algorithm. Subsequently, mammograms are classified by comparing textural and morphological features extracted from each cluster

Abstract

Automatic classification of breast density according BIRADS categories using a clustering approach

Methodology

Results

Reyer ZwiggelaarUniversity of Wales

rrz@aber.ac.uk

Josep PontJosep Trueta Hospital

rd.jpont@htrueta.scs.es

1) Breast SegmentationTwo-step algorithm: an adaptive threshold extracts the background and the annotations, while the pectoral muscle is segmented by using a region growing algorithm

2) Cluster those pixels with similar appearanceWe used Fuzzy C-Means with two seeds depicting fatty and dense tissue. Both seeds were initialized by using histogram information

3) Extract features for both clustersWe extracted two different sets of features: morphologic set (centre of masses, relative area, mean intensity) and textural set (derived from co-occurrence matrices)

4) Training the classifiersWe used two different classifiers: the k-Nearest Neighbours and the ID3 Decision Tree

68

1474B-I

17322522

B-III

121791764B-IV1728176284B-III10153982618B-II91332123757B-I

B-IVB-IIIB-IIB-IVB-IIB-IID3kNN

40159Low

7268

Dense

7442Dense47138Low

DenseLowID3kNN

We have proposed a novel approach to classify mammograms according their internal tissue, and into BIRADS categories. We tested it using MIAS and DDSM databases, obtaining an overall percentage of correct classification of 66% and 73% for kNN and ID3 classifiers when testing the MIAS database, and almost 80% for the Fuzzy combination when testing the DDSM database

Conclusions

Leave-one-woman-outWe used a leave-one-out procedure: each mammogram of the set was classified by using all the other mammograms except the opposite one (the other mammogram of the same woman) in the training setThe method was tested using 300 mammograms of the DDSM database, as well as 320 mammograms of the MIAS database, which were classified in BIRADS categories by an expert

Constructing the data

Testing the model

9183424B-I

2354305

B-III

171951593B-IV25401718253B-III7104983527B-II332042417B-I

B-IVB-IIIB-IIB-IVB-IIB-IID3kNN

49127Low

11047

Dense

10140Dense23103Low

DenseLowID3kNN

18207523522238124832332123

B-IVB-IIIB-IIB-IFc (kNN,ID3)

1133726124

DenseLowFc (kNN,ID3)

Test Image

Database of mammograms already classified

Tissue Classification

BIRADS

Training Step

Classification Step

Classifier

or

ID3

kNN

Classifier Approach

MIA

S

DD

SM

Feat_C1 Feat_C2

0) Database of already classified mammogramsThe method needs as starting point a subset of previously classified mammograms

3( ) ( ) (1 ) ( )FzC i kNN i ID iP mammo B P mammo B P mammo Bα α∈ = ∈ + − ∈1( )

( , )i

kNN ij Bj kNN

P mammo Bdist j mammo∈

∈

∈ = ∑

3

( )( , )

kj

k Nodes ID mammo

wP mammo leafdist feat feat

∈ = ∑

3( ) ( )i

ID i jj B

P mammo B P mammo leaf∈

∈ = ∈∑

( ) ( )

5) Fuzzy combination of both classifiersIn some cases, both classifiers give us complementary results. Hence, we constructed a new classifier which takes advantage of both approaches by converting this classifier approach into a probabilistic framework