Classification of boar sperm head images using Learning

Classification of boar sperm head images using Learning Vector Quantization Lidia Sánchez Michael Biehl, Piter Pasma, Marten Pijl, Nicolai Petkov University of León / Spain Rijksuniversiteit Groningen/ NL Electrical and Electronical Engineering Mathematics and Computing Science http: //www. cs. rug. nl/~biehl m. biehl@rug. nl

Motivation semen fertility assessment: important problem in human / veterinary medicine medical diagnosis: - sophisticated techniques, e. g. staining methods - high accurracy determination of fertility evaluation of sample quality for animal breeding purposes - fast and cheap method of inspection here: - microscopic images of boar sperm heads (Leon/Spain) e. g. quality inspection after freezing and storage - distance-based classification, parameterized by prototypes - Learning Vector Quantization + Relevance Learning ESANN 2006, Classification of boar sperm head images using LVQ

microscopic images of boar sperms preprocessing: - isolate and align head images - normalize with respect to mean grey level and corresponding variance - resize and approximate by an ellipsoidal region of 19 x 35 pixels - replace “missing” pixels (black) by the overall mean grey level ESANN 2006, Classification of boar sperm head images using LVQ

1360 example images, classified by experts (visual inspection) normal non-normal (650) (710) application of Learning Vector Quantization: - prototypes determined from example data parameterize a distance based classification plausible, straightforward to interpret/discuss with experts include adaptive metrics in relevance learning ESANN 2006, Classification of boar sperm head images using LVQ

Learning Vector Quantization (LVQ) example: basic scheme LVQ 1 [Kohonen] • initialize prototype vectors for different classes • present a single example classification: assignment ofclosest a vector • identify the prototype, the so-called winner toi. e the class of the closest prototype w • move the winner - closer towards the data (same class) aim: generalization ability(different class) - away from the data classification of novel data after learning from examples ESANN 2006, Classification of boar sperm head images using LVQ

Learning algorithms LVQ 1 Euclidean distance between data ξ prototype w: given ξ, update only the winner: (sign acc. to class membership) prototype initialization: class-conditional means + random displacement (∼ 70% correct classification) decreasing learning rate : ESANN 2006, Classification of boar sperm head images using LVQ

example outcome: LVQ 1 with 4 prototypes for each class: normal non-normal cross-validation scheme evaluation of performance - with respect to the training data, e. g. - with respect to test data average outcome over 90% of all data 10 realizations ESANN 2006, Classification of boar sperm head images using LVQ

ten-fold cross-validation: comparison of different LVQ systems (# of prototypes) performance on training data correct performance w. r. t. test data correct % al non-norm … improves with increasing number of (non-normal) prototypes al al rm rm no no % al non-norm … depends only weakly on the considered number of prototypes ESANN 2006, Classification of boar sperm head images using LVQ

Generalized Learning Vector Quantization (GLVQ) [A. S. Sato and K. Yamada, NIPS 7, 1995)] given a single example, update the two winning prototypes : w. J from the same class as the example (correct winner) w. K from the other class (wrong winner) perform gradient descent steps with respect to an instantaneous cost function f(z) ESANN 2006, Classification of boar sperm head images using LVQ

Generalized Relevance LVQ (GRLVQ) [B. Hammer, T. Villmann, Neural Networks 15: 1059 -1068] GLVQ with modified distance measure vector of relevances, normalization - re-define cost function f(z) in terms of dλ: - perform gradient steps w. r. t. prototypes w. J , w. K and vector λ GRLVQ - determines favorable positions of the prototypes - adapts the corresponding distance measure ESANN 2006, Classification of boar sperm head images using LVQ

Comparison of performance: estimated test error normal/non-normal prototypes alg. 3/3 1/7 LVQ 1 81. 4 % (4. 0) 81. 6 % (4. 5) GLVQ 75. 6 % (4. 1) 76. 4 % (3. 8) GRLVQ 81. 5 % (3. 5) 81. 7 % (3. 7) mean (stand. dev. ) - weak dependence on the number of prototypes - inferior performance of GLVQ (cost function ↮ classification error) - recovered when including relevances ESANN 2006, Classification of boar sperm head images using LVQ

GRLVQ: resulting relevances - only very few pixels are sufficient for successful classification test error: (all) 82. 75%, (69) 82. 75%, (15) 81. 87% normal (LVQ 1 prototypes) non-normal ESANN 2006, Classification of boar sperm head images using LVQ

Summary LVQ provides a transparent, plausible classification of microscopic boar sperm head images Performance: LVQ 1 ↘ GLVQ ↗ GRLVQ satisfactory classification error (ultimate goal: estimation of sample composition) Relevances: very few relevant pixels, robust performance noisy labels / insufficient resolution? Outlook - improve LVQ system, algorithms, relevance schemes - training data, objective classification (staining method) - classification based on contour information (gradient profile) ESANN 2006, Classification of boar sperm head images using LVQ

LVQ 1 demo ESANN 2006, Classification of boar sperm head images using LVQ