Deepening into the suitability of using pre-trained models of ImageNet against a lightweight convolutional neural network in medical imaging: an experimental study

Abstract

Transfer learning (TL) has been widely utilized to address the lack of training data for deep learning models. Specifically, one of the most popular uses of TL has been for the pre-trained models of the ImageNet dataset. Nevertheless, although these pre-trained models have shown an effective performance in several domains of application, those models may not offer significant benefits in all instances when dealing with medical imaging scenarios. Such models were designed to classify a thousand classes of natural images. There are fundamental differences between these models and those dealing with medical imaging tasks regarding learned features. Most medical imaging applications range from two to ten different classes, where we suspect that it would not be necessary to employ deeper learning models. This paper investigates such a hypothesis and develops an experimental study to examine the corresponding conclusions about this issue. The lightweight convolutional neural network (CNN) model and the pre-trained models have been evaluated using three different medical imaging datasets. We have trained the lightweight CNN model and the pre-trained models with two scenarios which are with a small number of images once and a large number of images once again. Surprisingly, it has been found that the lightweight model trained from scratch achieved a more competitive performance when compared to the pre-trained model. More importantly, the lightweight CNN model can be successfully trained and tested using basic computational tools and provide high-quality results, specifically when using medical imaging datasets.

A Two-Stage Model for Predicting Mild Cognitive Impairment to Alzheimer’s Disease Conversion

“…The performance of transfer learning is affected by the various factors such as the size of pretrained samples, the relevance of the source and target domains. Thus, not all the transfer learning can improve the model’s performance ( Huh et al, 2016 ; Zhuang X. et al, 2019 ; Alzubaidi et al, 2020 , 2021 ; Mustafa et al, 2021 ). For example, Alzubaidi et al (2021) found that the model trained from scratch performed better than those pretrained by ImageNet using three different medical imaging datasets.…”

“…Thus, not all the transfer learning can improve the model’s performance ( Huh et al, 2016 ; Zhuang X. et al, 2019 ; Alzubaidi et al, 2020 , 2021 ; Mustafa et al, 2021 ). For example, Alzubaidi et al (2021) found that the model trained from scratch performed better than those pretrained by ImageNet using three different medical imaging datasets. This observation inspired us to develop a two-stage model.…”

Brain Tumor Characterization Using Radiogenomics in Artificial Intelligence Framework

“…The well-known DL models such as convolutional neural network (CNN) and recurrent neural network (RNN) have been successfully used in computer vision studies. The pre-trained model of CNN, such as AlexNet, ResNet, GoogleNet, etc., training on the voluminous ImageNet dataset, showed quite good performance of image features of the multi-set datasets [ 125 ]. The transfer learning (TL) also adds dimensionality to the deep feature extraction and successful computation with a suitable classifier.…”

Face Recognition Based on Deep Learning and FPGA for Ethnicity Identification

“…Recently, researchers have attracted significant attention to Deep Learning (DL) [11][12][13][14] owing to its numerous applications in speech processing [15], natural language processing [16], and CV [17,18]. In video recognition [19] and large-scale images, a model of DL so-called convolutional neural network (CNN) has lately attained several encouraging results.…”