Institute for Systems and Robotics Instituto de Medicina Molecular Automatic Pairing of Metaphase Chromosomes with Mutual Information for Karyotyping Purposes Artem Khmelinskii 1, Rodrigo Ventura 1, Carmo-Fonseca, M. 2 and João Sanches Systems and Robotic Institute / Instituto Superior Técnico 2 Instituto de Medicina Molecular / Faculdade de Medicina de Lisboa Lisboa, Abstract Clinical cytogenetics plays a key role in the diagnosis of genetic diseases and haematological disorders that involve chromosomal abnormalities. A major technique used in cytogenetics is karyotyping. A karyotype analysis consists of blocking cells in mitosis and staining the chromosomes. The 46 human chromosomes are then paired and visually arranged in decreasing order of size.The pairing procedure aims at to identify all pairs of homologous chromosomes. The pairing criterion is based on dimensional, morphological, and textural features similarity. This process is time consuming when performed manually, and demanding from a technical point of view. An automatic pairing algorithm would thus bring benefits, but it remains an open problem to date. In this work 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. A supervised non-linear classifier is trained by using manual classifications provided by expert technicians, combining the different features computed from each pair. Simulations using 836 real chromosome images taken from bone marrow samples of patients with haematological disorders, obtained with a Leica Optical Microscope DM 2500, in a leave-one-out cross-validation fashion, were performed for training and testing the algorithm. Furthermore, qualitative comparisons with the results obtained with the Leica CW 4000 Karyo software were also performed. Promising and relevant results were found, despite the poor quality of the original chromosome images, contrasting with state-of-the-art algorithms and datasets found in the literature. Feature Extraction 1.Size – Area, Perimeter, Bounding Box dimensions, Aspect Ratio 2.Shape – Normalized area (Perimeter/Area) 3.Texture – Mutual Information (pair), Band Profile PairsF1F1 F2F2...FLFL P 11 d 1 (1,1)d 2 (1,1)...d L (1,1) P 12 d 1 (1,2)d 2 (1,2)...d L (1,2)... P NN d 1 (N,N)d 2 (N,N)...d L (N,N) Data – Feature/Distance Matrix Pre-processing 1.Metaphase Chromosome extraction 2.Geometrical Compensation 3.Geometrical Scaling 4. Intensity Compensation p i,j c1c1 c2c2...c N-1 cNcN c1c c2c c N cNcN Classifier  Binary Global Optimization Find a permutation matrix,, where Mutual Information Withoutwith Data Set 194.4% Data Set 293.5%95.4% Data Set 368.9%70.1% Training 1.Distance between two chromosomes p and q: 2.Training -- computing w r Results This work was partially supported by FCT (ISR/IST plurianual funding) through the POS_Conhecimento Program that includes FEDER funds.