Table 3.
A list of recent papers related to medical image detection
| Author | Year | Application | Model | Dataset | Contributions highlights |
|---|---|---|---|---|---|
| Specific-type medical objects detection | |||||
| Ding et al., 2017 | 2017 | Lung nodules detection from CT images | Faster RCNN with changed VGG16 as backbone | LUNA16 | (1) Using deconvolutional layer to recover fine-grained features; (2) Using 3D CNN to exploit 3D spatial context information for false positives reduction. |
| Zhu et al., 2018 | 2018 | Lung nodules detection from CT images | 3D Faster RCNN with U-Net-like structure, built with dual path blocks | LIDC-IDRIs | (1) Using 3D Faster RCNN considering the 3D nature of lung CT images; (2) Utilizing the compactness (i.e., fewer parameters) of dual path networks on small dataset. |
| Wang et al., 2020c | 2020 | Lung nodules detection from CT images | 3D variant of FPN with modified residual network as backbone | LUNA16 and NLST | (1) A semi-supervised learning strategy to leverage unlabeled images in NLST; (2) Mixup augmentation for examples with pseudo labels and ground truth annotations; (3) FPN outputs multi-level features to enhance small object detection. |
| Mei et al., 2021 | 2021 | Lung nodules detection from CT images | U-shaped architecture, with 3D ResNet50 as encoder | PN9 | (1) Inserting non-local modules in residual blocks to seize long-range dependencies of different positions and different channels. (2) Using multi-scale features for false positives reduction. |
| Ma et al., 2021a | 2020 | Breast mass detection from mammograms | CVR-RCNN: Two-branch Faster RCNNs, with relation modules (Hu et al., 2018b) | DDSM and a private dataset | Extraction of complementary relation features on CC and MLO views of mammograms using relation modules. |
| Liu et al., 2020c | 2020 | Breast mass detection from mammograms | BG-RCNN: Incorporating Bipartite Graph convolutional Network (BGN) into Mask RCNN | DDSM and a private dataset | (1) Modeling relations (e.g., complementary information and visual correspondences) between CC and MLO views of mammograms using BGN; (2) Defining simple pseudo landmarks in mammograms to facilitate learning geometric relations. |
| Rijthoven et al., 2018 | 2018 | Lymphocytes detection in whole-slide (WSI) histology images of breast, colon, and prostate cancer | Smaller YOLOv2 with much fewer layers | Private dataset | (1) Simplifying the original YOLO network using prior knowledge of lymphocytes (e.g., average size, no overlaps); (2) Designing a new training sampling strategy using the prior knowledge (i.e., brown areas without lymphocytes contain hard negative samples). |
| Lin et al., 2019 | 2019 | Lymph node metastasis detection from WSI histology images | Modified Fully convolutional network (FCN) based on VGG16 | Camelyon16 dataset and ISBI 2016 | (1) Utilizing FCN for fast gigapixel-level WSI analysis; (2) Proposing anchor layers for model conversion to ensure dense scanning; (3) Hard negative mining. |
| Nair et al., 2020 | 2020 | Multiple sclerosis lesion detection from MR brain images | 3D U-Net based segmentation network to obtain lesions | Private dataset | (1) Uncertainty estimation using Monte Carlo (MC) dropout; (2) Using multiple uncertainty measures to filter out uncertain predictions of lesion candidates. |
| Universal lesion detection | |||||
| Yan et al., 2018a | 2018 | Detection of lung, mediastinum, liver, soft tissue, pelvis, abdomen, kidney, and bone lesions from CT images | 3DCE: Modified R-FCN | DeepLesion | (1) Exploiting 3D context information; (2) Leveraging pre-trained 2D backbones (VGG-16) for transfer learning. |
| Tang et al., 2019 | 2019 | Detection of various types of lesions in DeepLesion | ULDor: Mask RCNN with ResNet-101 as backbone | DeepLesion | (1) Pseudo mask construction using RECIST annotations; (2) Hard negative mining to learn more discriminative features for false positives reduction. |
| Yan et al., 2019 | 2019 | Detection of various types of lesions in DeepLesion | MULAN: Modified Mask RCNN with DenseNet-121 as backbone | DeepLesion | (1) Jointly performing three different tasks (detection, tagging, and segmentation) for better performance; (2) A new 3D feature fusion strategy. |
| Tao et al., 2019 | 2019 | Detection of various types of lesions in DeepLesion | Improved R-FCN | DeepLesion | Contextual attention module aggregates relevant context features, and spatial attention module highlights discriminative features for small objects. |
| Li et al., 2019 | 2019 | Detection of various types of lesions in DeepLesion | MVP-Net: a three pathway architecture with FPN as backbone | DeepLesion | Using an attention module to incorporate clinical knowledge of multi-view window inspection and position information. |
| Unsupervised lesion detection | |||||
| Baur et al., 2021 | 2021 | Segmentation/detection of brain MRI | A collection of VAE- and GAN-based models | Private data, MSLUB, MSSEG2015 | A comprehensive and in-depth investigation into the strengths and shortcomings of a variety of methods for anomaly segmentation. |
| Chen et al., 2021c | 2021 | Detection of MRI brain tumors and stroke lesions | VAE-based model | CamCAN, BRATS17, ATLAS | Proposing a more accurate approximation of VAE’s original loss by replacing the gradients of ELBO with the derivatives of local Gaussian distributions. |
| Chen et al., 2020d | 2020 | MRI glioma and stroke detection | VAE-based model | CamCAN, BRATS17, ATLAS | Using autoencoding-based methods to learn a prior for healthy images and using MAP estimation to for image restoration. |
| Schlegl et al., 2017 | 2017 | Anomaly detection in optical coherence tomography (OCT) | AnoGAN: DCGAN-based model | Private dataset | (1) The first work using GAN for anomaly detection; (2) Proposing a new approach that iteratively maps input images back to optimal latent representations for anomaly detection. |
| Schlegl et al., 2019 | 2019 | OCT anomaly detection | WGAN-based model | Private dataset | Based on AnoGAN, an additional encoder was introduced to perform fast inverse mapping from image space to latent space. |
| Baur et al., 2018 | 2018 | MRI multiple sclerosis detection | AnoVAEGAN: a combination of VAE and GAN | Private dataset | (1) Combining VAE and GAN for fast inverse mapping; (2) The model can operate on an entire MR slice to exploit global context. |
| Uzunova et al., 2019 | 2019 | MRI brain tumor detection | CVAE-based model | BRATS15 | Utilizing location-related condition to provide additional prior information of healthy and unhealthy tissues for better performance. |