Merging databases for CNN image recognition, increasing bias or improving results?

Abstract

Automated microscopy, image processing, and recognition using artificial intelligence is getting a growing interest from the scientific community, as more and more research centres are actively working on building datasets of images for training convolutional neural networks (CNNs) to identify microscopic objects. However, images acquired between institutes might show differences in light and contrast intensity leading to potential bias in identification when using datasets or CNNs from another institute.

One might then question if combining datasets acquired in different conditions is likely to improve the efficiency of the resulting CNN by increasing the number of images and integrating lighting variability into the learning process, or on the contrary introduce bias in the CNN training by adding a recurrent noise, common to all classes, through a substantial light and contrast variability.

In order to ease collaboration between laboratories, two datasets of middle Eocene radiolarian images, acquired separately at GNS Science (NZ) and the University of Lille (France), were generated to assess the accuracy of CNNs trained on both datasets individually, and also when combined into a third dataset. The performance of the three resulting CNNs was then evaluated on new images acquired at both institutions.

Finally, the new radiolarian dataset generated at GNS allowed to easily detect unknown taxa, which are otherwise abundant in the studied material. Seven new species are described: Ceratospyris metroid n. sp., Ceratospyris okazakii n. sp., Desmospyris biloba n. sp., Botryostrobus lagena n. sp., Buryella apiculata n. sp., Lophocyrtis cortesei n. sp., and Cromyosphaera fulgurans n. sp.