Diapositivo 1

1
Chromosome Pairing Using Mutual Information in
Bone Marrow Cells
Artem Khmelinskii Rodrigo
Ventura João Sanches (artkhmelinskii_at_isr.ist
.utl.pt) (rodrigo.ventura_at_isr.ist.utl.pt)
(jmrs_at_ist.utl.pt) Institute for Systems and
Robotics / Instituto Superior Técnico1049-001
Lisbon, Portugal
Learning the weights For each chromosome type, a
weight vector is learnt from a training set of
manually classified karyotypes. For each
chromosome type, the corresponding weight vector
is obtained by minimizing the energy
function with respect to the weight vector w
Abstract This paper is focused in the specific
task of pairing chromosomes from bone marrow
cells extracted from leukemia patients in the
scope of karyotyping procedure. In this paper a
new strategy for automatic pairing of homologous
chromosomes is proposed. Besides the traditional
features described in the literature, the Mutual
Information (MI) is used to discriminate
chromosome textural differences. The goal is to
better characterize the textural information
associated with each pair by adding
discriminative power to the G-banding profiles
information.
Pairing Given a distance matrix D, finding the
optimal pairing that minimizes the sum of the
distances between each chromosome pair can be
cast as an integer programming problem with the
following formulation

• Problem Formulation
• Given a karyotype, and assuming that all
  chromosomes are already segmented, a distance
  matrix D is computed, where each entry D(i,j)
  represents the distance between chromosomes i and
  j. This distance is computed by minimizing
  weighted sums of a set of mutual features, over a
  set of weight vectors.
• Before any feature extraction, all chromosomes
  are subject to a pre-processing step, consisting
  of bounding box detection, geometric correction,
  edge regularization, and histogram equalization.
• One of the mutual features is the mutual
  information (MI) between two given chromosomes
• All of the other features correspond to Euclidean
  distances between features of each one of the
  chromosomes. These individual features are
• size area, perimeter, height, length of minor
  axis, and length proportion
• shape normalized area (area/perimeter ratio)
• band pattern density profile

Experimental Results Three datasets, for which
ground truth manual classification was provided,
were used. The following table contains
the results obtained, in the form of the average
successful pairing, using a leave-one-out cross
validation (LOOCV) technique.
Data set 1 (N4)
(all) Data set 3 (N22)
Data set 2 (N8)
Conclusions In this work a new pairing algorithm
for bone marrow chromosomes is proposed without
chromosome classification. The features,
associated with each pair, are used by a battery
of 22 classifiers. Additionally, the Mutual
Information is also proposed as a new feature to
increase the discriminative power of the
algorithm w.r.t. the band profile of the
chromosomes.