Removing Zero Variance Units of Deep Models for COVID-19 Detection

Abstract

Deep Learning has been used for several applications including the analysis of medical images. Some transfer learning works show that an improvement in performance is obtained if a pre-trained model on ImageNet is transferred to a new task. Taking into account this, we propose a method that uses a pre-trained model on ImageNet to fine-tune it for Covid-19 detection. After the fine-tuning process, the units that produce a variance equal to zero are removed from the model. Finally, we test the features of the penultimate layer in different classifiers removing those that are less important according to the f-test. The results produce models with fewer units than the transferred model. Also, we study the attention of the neural network for classification. Noise and metadata printed in medical images can bias the performance of the neural network and it obtains poor performance when the model is tested on new data. We study the bias of medical images when raw and masked images are used for training deep models using a transfer learning strategy. Additionally, we test the performance on novel data in both models: raw and masked data.

Keywords

• Covid-19 detection
• model compression
• transfer learning
• deep learning explainability