Are Convolutional Neural Networks Trained on ImageNet Images Wearing Rose‐Colored Glasses?: A Quantitative Comparison of ImageNet, Computed Tomographic, Magnetic Resonance, Chest X‐Ray, and Point‐of‐Care Ultrasound Images for Quality

Abstract

Objectives Deep learning for medical imaging analysis uses convolutional neural networks pretrained on ImageNet (Stanford Vision Lab, Stanford, CA). Little is known about how such color‐ and scene‐rich standard training images compare quantitatively to medical images. We sought to quantitatively compare ImageNet images to point‐of‐care ultrasound (POCUS), computed tomographic (CT), magnetic resonance (MR), and chest x‐ray (CXR) images. Methods Using a quantitative image quality assessment technique (Blind/Referenceless Image Spatial Quality Evaluator), we compared images based on pixel complexity, relationships, variation, and distinguishing features. We compared 5500 ImageNet images to 2700 CXR, 2300 CT, 1800 MR, and 18,000 POCUS images. Image quality results ranged from 0 to 100 (worst). A 1‐way analysis of variance was performed, and the standardized mean‐difference effect size value (d) was calculated. Results ImageNet images showed the best image quality rating of 21.7 (95% confidence interval [CI], 0.41) except for CXR at 13.2 (95% CI, 0.28), followed by CT at 35.1 (95% CI, 0.79), MR at 31.6 (95% CI, 0.75), and POCUS at 56.6 (95% CI, 0.21). The differences between ImageNet and all of the medical images were statistically significant (P ≤ .000001). The greatest difference in image quality was between ImageNet and POCUS (d = 2.38). Conclusions Point‐of‐care ultrasound (US) quality is significantly different from that of ImageNet and other medical images. This brings considerable implications for convolutional neural network training with medical images for various applications, which may be even more significant in the case of US images. Ultrasound deep learning developers should consider pretraining networks from scratch on US images, as training techniques used for CT, CXR, and MR images may not apply to US.

Simplified Lung Ultrasound Examination and Telehealth Feasibility in Early SARS-CoV-2 Infection

“…Spatial and temporal resolution and deep learning algorithms have demonstrated variable effects in early studies, having the potential for greater or lesser sensitivity, and are likely bounded by the subjective image search and acquisition process. [45][46][47] Nonetheless, this study has demonstrated that a specific ultrasound examination can be performed during telehealth by patients without any prior training and has future implications for this methodology in other disease states. In COVID-19, the simplified lung examination could be performed not only from home isolation by patients but also in emergency departments, urgent care facilities, or COVID-19 testing centers for outpatients with minimal symptoms.…”

Discriminative deep learning based benignity/malignancy diagnosis of dermatologic ultrasound skin lesions with pretrained artificial intelligence architecture

“…Large image databases such as Imagenet 12 permit initial weights preadjustment in order to make borders or other basic features of images more easily recognizable by algorithms 20 . However, in the case of ultrasound, some authors suggest that initial weight adjustments should be made with the same kind of images to improve system accuracy 21 . This is a controversy, which could foster the necessity of creating largely public ultrasound images databases.…”

Making Artificial Intelligence Lemonade Out of Data Lemons

“…For algorithm creation, researchers used an Anaconda package manager with Python programming language version 3.72, to facilitate scripting and package management. Based on superior performance in prior studies, we decided upon a publicly available Keras‐based (a python deep learning framework or library) VGG‐16 bidirectional LSTM ML algorithm 18 . VGG‐16 model architecture is accessible from multiple public sources including an online repository, http://github.com.…”

“…Based on superior performance in prior studies, we decided upon a publicly available Kerasbased (a python deep learning framework or library) VGG-16 bidirectional LSTM ML algorithm. 18 VGG-16 model architecture is accessible from multiple public sources including an online repository, github.com. The VGG convolutional neural network (CNN) is an early CNN version and contains 16 layers, while most recent CNNs have hundreds of layers.…”