Table 1.

A list of recent papers related to medical image classification

Author	Year	Application	Model	Dataset	Contributions highlights
Supervised classification
Guan et al., 2018	2018	Thorax disease classification from chest X-rays	AG-CNN: an attention guided CNN	ChestX-ray14	(1) Using attention mechanism to identify discriminative regions from the global image, which were used to train the local CNN branch; (2) Fusing the local and global information for better performance.
Schlemper et al., 2019	2019	2D fetal ultrasound image plane classification	AG-Sononet: attention-gated model	Private dataset	Incorporating grid attention into Sononet (Baumgartner et al., 2017) to better exploit local information and aggregating attention vectors at different scales for final prediction.
Unsupervised image synthesis
Frid-Adar et al., 2018b	2018	CT liver lesion classification	DCGAN	Private dataset	Using the high-quality liver ROIs synthesized by DCGAN to perform data augmentation.
Frid-Adar et al., 2018a	2018	CT liver lesion classification	ACGAN	Private dataset	Comparing GAN’s augmentation performance in conditional and unconditional settings.
Wu et al., 2018a	2018	Mammogram classification	cGAN	DDSM dataset	Controlling generating a specific type of lesions using malignant/non-malignant labels.
Self-supervised learning based classification
Azizi et al., 2021	2021	Classification of chest X-ray images and dermatology images	MICLe: based on SimCLR (Chen et al., 2020a)	CheXpert, and private dataset	Proposing a new contrastive learning approach based on SimCLR by leveraging multiple images of each medical condition for additional self-supervised pretraining.
Vu et al., 2021	2021	Classification of pleural effusion in chest X-ray images	MedAug: Based on MoCo (Chen et al., 2020b)	CheXpert	(1) Utilizing patient metadata to create positive pairs for contrastive learning; (2) Showing that self-supervised pre-training can perform better than ImageNet pre-training.
Chen et al., 2021a	2021	COVID-19 diagnosis from chest CT images	MoCo-based classification	DeepLesion, LIDC-IDRI, UCSD COVID-19 CT, SIRM’s COVID-19 data	Using contrastive learning to pre-train an encoder on public datasets so that expressive features of non-COVID CT images can be captured, and using the pre-trained encoder for few-shot COVID-19 classification.
Sowrirajan et al., 2021	2021	Classification of pleural effusion and tuberculosis from chest X-rays	MoCo-CXR: Based on MoCo (Chen et al., 2020b)	CheXpert, Shenzen dataset	Showing that MoCo-pretrained feature representations on large X-ray databases can (1) outperform ImageNet pre-training on downstream tasks with small, labeled X-rays, and (2) generalize well to an external dataset.
Chen et al., 2019b	2019	Fetal ultrasound image plane classification	A general CNN-based architecture	Private dataset	Designing a new self-supervised pretext task based on context restoration to learn high-quality features from unlabeled images.
Zhou et al., 2021	2021	CT lung nodule false positive reduction, etc.	Models Genesis	LUNA 2016, etc.	Consolidating four different self-supervised schemes (non-linear, local-shuffling, inner and outer cutouts) to learn representations from different perspectives (appearance, texture, and context).
Semi-supervised learning based classification
Liu et al., 2020a	2020	Thorax disease classification from chest X-rays, etc.	Based on Mean Teacher (Tarvainen and Valpola, 2017)	ChestX-ray14, etc.	Proposing sample relation consistency for the semi-supervised model to extract useful semantic information from the unlabeled data.
Xie et al., 2019a	2019	CT lung nodule classification	Adversarial autoencoder-based model	LIDC-IDRI, Tianchi Lung Nodule dataset	Using learnable transition layers to enable transferring representations from the reconstruction network to the classification network.
Madani et al., 2018a	2018	Cardiac abnormality classification in chest X-rays	Semi-supervised GAN	NIH PLCO dataset, etc.	Employing a semi-supervised GAN architecture to address the scarcity of labeled data.