Abstract
Subtle morphological changes caused by toxic chemicals in different regions of organs in small zebrafish embryos can be difficult to detect by human eyes or even by traditional image processing tools. Without brute force testing arbitrary parameters and thresholds in various combinations of imaging algorithms, we present our systematic approach in which we first train convolutional neural networks (CNNs) to recognize which organs of which zebrafish embryos contain significant morphological differences so they can be reliably classified as being exposed to either the fresh water or the toxic water contaminated with chemicals. From the training results, the organ images in each identified image subset were then analyzed by the technique, occlusion sensitivity map, to allocate the exact locations and the scales of morphological changes caused by different chemicals. In the experiment section of this paper, we use multiple chemical exposure scenarios to demonstrate the morphological differences in organs that can be detected by our approach.