A scoping review of transfer learning research on medical image analysis using ImageNet

Morid, Mohammad Amin; Borjali, Alireza; Fiol, Guilherme Del

doi:10.1016/j.compbiomed.2020.104115

Cited by 299 publications

(179 citation statements)

References 81 publications

Supporting

Mentioning

144

Contrasting

Unclassified

Order By: Relevance

“…This would lead to network with poor performance if the training was simply performed on the small amount of data. Transfer learning [10], which starts with a model pretrained with a large dataset, is a popular method to address this issue, as a better starting point for the model generally will lead to better results.…”

Section: Introductionmentioning

confidence: 99%

3D Attention U-Net with Pretraining: A Solution to CADA-Aneurysm Segmentation Challenge

Jia²,

Liao

et al. 2021

Cerebral Aneurysm Detection

View full text Add to dashboard Cite

Early detection and accurate segmentation of cerebral aneurysm is important for clinical diagnosis and prevention of rupture, which would be life threatening. 3D images can provide abundant information for characterizing the aneurysm. But the traditional manual segmentation of aneurysms takes lots of time and effort. Therefore, accurate and rapid automatic algorithm for 3D segmentation of aneurysm is needed. U-Net is a widely used deep learning network in medical image segmentation, but its performance is limited by the amount of data. In this challenge of aneurysm segmentation, we proposed to add attention gate and Models Genesis pretraining mechanisms to the classical U-Net model to improve the results. The dice of 3D U-net, 3D Attention U-Net, pretrained 3D U-Net and pretrained 3D Attention U-Net are 0.881, 0.884, 0.890 and 0.907, respectively. The experimental results show that the use of attention gate and Models Genesis can significantly improve the performance of U-Net model in segmenting aneurysms. This work achieved rank one in CADA 2020-Aneurysm Segmentation Challenge.

show abstract

Section: Introductionmentioning

confidence: 99%

3D Attention U-Net with Pretraining: A Solution to CADA-Aneurysm Segmentation Challenge

Jia²,

Liao

et al. 2021

Cerebral Aneurysm Detection

View full text Add to dashboard Cite

show abstract

“…A fine-tuning approach, such as that of the feature extractor, utilizes a well-trained CNN model on a large dataset, such as ImageNet, as the base and supersedes the CNN layers with new CNN layers [73,74]. In fine-tuning, instead of freezing the convolution layers of the well-trained CNN model, their weights are updated during the training process [51][52][53].…”

Section: Fine-tuningmentioning

confidence: 99%

“…The main advantage of the CNN is its accuracy in image recognition; however, it involves a high computational cost and requires numerous training data [85]. A CNN generally comprises an input layer, one or many convolution layers, pooling layers, and a fully connected layer [74]. The following are the most commonly used CNN models used for transfer learning with breast ultrasound images [84].…”

Section: Convolutional Neural Networkmentioning

confidence: 99%

Transfer Learning in Breast Cancer Diagnoses via Ultrasound Imaging

Ayana

Dese

Choe

2021

Cancers

103

View full text Add to dashboard Cite

Transfer learning is a machine learning approach that reuses a learning method developed for a task as the starting point for a model on a target task. The goal of transfer learning is to improve performance of target learners by transferring the knowledge contained in other (but related) source domains. As a result, the need for large numbers of target-domain data is lowered for constructing target learners. Due to this immense property, transfer learning techniques are frequently used in ultrasound breast cancer image analyses. In this review, we focus on transfer learning methods applied on ultrasound breast image classification and detection from the perspective of transfer learning approaches, pre-processing, pre-training models, and convolutional neural network (CNN) models. Finally, comparison of different works is carried out, and challenges—as well as outlooks—are discussed.

show abstract

“…16 We initialized EfficientNet-B4 with the parameter values pretrained on ImageNet and continued training the entire model, not just the last few fully-connected layers. 3 Combining and then jointly training a small dataset of interest with a larger auxiliary dataset often help the prediction accuracy. 17,18 For the auxiliary dataset, we downloaded the publicly available SIIM-ISIC Melanoma Classification Challenge Dataset from 2018 to 2020.…”

Section: Classifiermentioning

confidence: 99%

“…For this approach to work well, the pretrained data must either be similar in type to the medical data or the size of the medical dataset must be relatively large. 3 Due to these limitations, neural network applications in healthcare have focused on relatively common conditions, where sufficiently large datasets are more readily collected. Genetic conditions, though common in aggregate, are largely individually rare.…”

Section: Introductionmentioning

confidence: 99%

Proof-of-principle neural network models for classification, attribution, creation, style-mixing, and morphing of image data for genetic conditions

Duong

Waikel

et al. 2021

Preprint

View full text Add to dashboard Cite

Neural networks have shown strong potential to aid the practice of healthcare. Mainly due to the need for large datasets, these applications have focused on common medical conditions, where much more data is typically available. Leveraging publicly available data, we trained a neural network classifier on images of rare genetic conditions with skin findings. We used approximately 100 images per condition to classify 6 different genetic conditions. Unlike other work related to these types of images, we analyzed both preprocessed images that were cropped to show only the skin lesions, as well as more complex images showing features such as the entire body segment, patient, and/or the background. The classifier construction process included attribution methods to visualize which pixels were most important for computer-based classification. Our classifier was significantly more accurate than pediatricians or medical geneticists for both types of images. Next, we trained two generative adversarial networks to generate new images. The first involved all of the genetic conditions and was used for style-mixing to demonstrate how the diversity of small datasets can be increased. The second focused on different disease stages for one condition and depicted how morphing can illustrate the disease progression of this condition. Overall, our findings show how computational techniques can be applied in multiple ways to small datasets to enhance the study of rare genetic diseases.

show abstract

A scoping review of transfer learning research on medical image analysis using ImageNet

Cited by 299 publications

References 81 publications

3D Attention U-Net with Pretraining: A Solution to CADA-Aneurysm Segmentation Challenge

3D Attention U-Net with Pretraining: A Solution to CADA-Aneurysm Segmentation Challenge

Transfer Learning in Breast Cancer Diagnoses via Ultrasound Imaging

Proof-of-principle neural network models for classification, attribution, creation, style-mixing, and morphing of image data for genetic conditions

Contact Info

Product

Resources

About