Deep Transfer Learning for Nucleus and Micronucleus Recognition

Nucleus is the main component in a human cell. Excellent nucleus shape observation methods help endorse scientific research in medical and biological fields. Nucleus shape observation methods are performed manually by humans through medical labs. Abnormal nucleus (Micronucleus) is caused by drugs and other toxical factors. Current methods are only confined in detecting and segmenting nucleuses images from different human tissues. None of the these methods has tackled the problem of automating the nucleus and micronucleus recognition. In this paper, first, we apply a deep transfer learning to automate the recognition of nucleus and micronucleus from microscopy images. Second, we introduce our dataset which consists of manually collected nucleus and micronucleus image examples to train and evaluate the performance of the proposed method. The examples are manually curated by domain of experts in the biology. The nucleus and micronucleus images are used as an input to feed several pre-trained Convolutional Neural Network (CNN) models. The models performance is measured using 2-fold cross validation. Our experimental evaluation shows that GoogLeNet pre-trained model performs better than other models with an 80.06 % of average classification accuracy.