Fast Convolutional Neural Network Training Using Selective Data Sampling Application to Haemorrhage Detection in Color Fundus Images.These models are based on convolution operations applied to the input image at multiple hierarchical layers. Recent works on cancer detection and brain segmentation have shown CNN achieved remarkable performance.

However, the need of large high-quality training sets to accurately train CNNs prevent a wider adoption of these networks in medical imaging.

In this work we focus on finding diseased regions in images, a common task in medical image analysis.

In such a classification task, CNNs are trained with small patches centered on pixels of interest. Treating uniformly this data during the learning process leads to many training iterations wasted on non-informative samples, making the CNN training process unnecessarily time-consuming.

We have presented a method to substantially speed-up the time-consuming training process of convolutional neural networks with a selective sampling strategy, named SeS, embedded in the training procedure. We have demonstrated excellent results in the identification of hemorrhages on color fundus images.