Computer Vision News - July 2016

things can get much more complicated, for instance when not all chromosomes can be seen clearly since during mitosis some chromosomes may overlap each other and some parts are hidden behind other parts. Also, this case is statistically quite particular, since we know in advance the number of chromosomes, which in normal human cells is 46. This problem calls for a linear 1d classifier that will take advantage of the fact the we are expecting a band- like appearance of the chromosome, which is a 1d function that we need to detect. On the other hand, we can also use a classical 2d Convolutional Neural Networks (CNN) to do the detection. As we all know, all projects in computer vision making use of deep learning will try as a first guess the “low-hanging fruit” approach: that means picking the most obvious and readily available opportunities. Instead of training and building a new classifier with specific structure, this approach (called Transfer Learning ) will use one of the known solutions (like ImageNet and others) and adapt it to our purpose. It will use the CNN part as feature detectors and use one or two of the upper layers as a Support Vector Machine (SVM) to train it to the specific classification. This approach usually gives much faster results to match limited resources and datasets. This transfer learning is also extremely precious for the developer in that it helps to overcome the lack of a large enough dataset and it gives appropriate results even in the many circumstances when access to a large and fully annotated database of cells and organs is not possible. Computer Vision News Project 39 Normal human male karyotype (complete set of chromosomes) “ Testing and classification of chromosomes plays a key role in our ability to understand genetic pathologies ” Project