Multiclass and Multilabel Classification of Human Cell Components Using Transfer Learning of InceptionV3 Model

Abstract

Here we are finding 28 different predefined proteins of human cells with a transfer learning of inceptionV3 as the base model. This base model is pre-trained on the imagenet dataset. We add some layers in the base model and trained resulting models for our dataset and find the human cell proteins in a given sample. For this, we used a human cell atlas dataset provided by the Human Protein Atlas (HPA) community. This is a multilabel and multiclass problem in which one sample image can have more than one protein. We trained the given model on the training dataset by using the best hyperparameter of deep learning then tested the trained model on test data. We have found the efficiency of the resulting model in terms of precision, recall, f1-score, micro average, macro average, weighted average, sampled average, and accuracy. The accuracy of the given resulted model is 95.96%. On the basis of the above parameters, we analyzed the performance of the mentioned model, hence we choose the best hyperparameter according to the performance matrices in case of multi labels and multiclass problems.