| Literature DB >> 36237706 |
Abstract
Deep learning has recently become one of the most actively researched technologies in the field of medical imaging. The availability of sufficient data and the latest advances in algorithms are important factors that influence the development of deep learning models. However, several other factors should be considered in developing an optimal generalized deep learning model. All the steps, including data collection, labeling, and pre-processing and model training, validation, and complexity can affect the performance of deep learning models. Therefore, appropriate optimization methods should be considered for each step during the development of a deep learning model. In this review, we discuss the important factors to be considered for the optimal development of deep learning models. CopyrightsEntities:
Year: 2020 PMID: 36237706 PMCID: PMC9431842 DOI: 10.3348/jksr.2020.0171
Source DB: PubMed Journal: Taehan Yongsang Uihakhoe Chi ISSN: 1738-2637
Fig. 1Visualization of deep learning models using Class Activation Map for colonoscopy images. Activation occurs at the location of the polyp.
Representative Open Data Sets for Medical Image Collection
| Provider | Data Name | Collection Type | Modality | Images |
|---|---|---|---|---|
| NCI | TCIA | Common disease (e.g. lung cancer, breast cancer, colorectal cancer, head and neck cancer) | CT, MR, MG, CR, DX, pathology, etc. | Over 100 unique data collections, 30 million images |
| NIH | NLST | Lung cancer | CT, pathology | 25000 patients, 200000 images |
| NIH | Deep lesion | Common disease (e.g. lung nodules, liver tumors, enlarged lymph nodes, etc.) | CT | 4427 patients, 32120 images |
| NIH | ChestX-ray8 | Common thoracic diseases (atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, pneumothorax) | X-ray | 32717 patients, 108948 images |
| Washington University Knight Alzheimer Disease Research Center | OASIS3 | Alzheimer disease | MRI, PET | 1098 patients, 2000 MR, 1500 PET |
| University of South Florida | DDSM | Benign and malignant mass | Mammography | 2620 images |
| University of Michigan, Deep Blue | RIGA | Glaucoma | Fundus image | 750 images |
CR = computed radiography, DX = digital X-ray, MG = mammography, NCI = National Cancer Institute, NIH = National Institutes of Health, NLST = The National Lung Screening Trial, TCIA = The Cancer Imaging Archive
Fig. 2Labeling using a pre-trained deep learning model based on the core dataset.
Free and Commercial Labeling Software for Medical Images
| Name | Region of Interest | URL | |
|---|---|---|---|
| Free software | |||
| ImageJ | Rectangle, oval, polygon, freehand |
| |
| ITK-snap | Polygon, paintbrush, smart brush |
| |
| Label me | Polygon, rectangle, circle, line and point |
| |
| CVAT | Box, polygon, points & line, point |
| |
| Medtagger | Slicing, brush, polygon, pointing |
| |
| Commercial software | |||
| Aview | Brush, polygons, magic cut |
| |
| MEDIP | Polygon, free-draw, pixel-wised, draw cut, region growing |
| |
| OsiriX | Rectangle, oval, polygon, point, brush, region growing |
| |
| Labelbox | Box, polygon, points & line, custom attributes |
| |
| Trainingdata | Segmentation with brush & eraser, bounding box, polygon tool, freehand with sculpter, key point tool, magnify, 2D growth tool for DICOM |
| |
Studies on Data Augmentation Methods
| Reference | Modality | Method | Techniques |
|---|---|---|---|
| Mehta & Arbel ( | MR | Standard | Translation, rotation, scaling, shear transformation |
| Isensee et al. ( | MR | Standard | Rotation, scaling, elastic deformation, gamma correction, mirroring |
| Rahman et al. ( | X-ray | Standard | Rotation, translation |
| Yang et al. ( | X-ray | Standard | Rotation, translation, zoom |
| Han et al. ( | MR | GAN | Conditional PGGAN |
| Sandfort et al. ( | CT | GAN | CycleGAN |
| Frid-Adar et al. ( | CT | GAN | Deep convolutional GAN, auxiliary classifier GAN |
| Xin et al. ( | MR | GAN | Pix2pix |
GAN = Generative Adversarial Network
Fig. 3Example of CT image pre-processing.
A. Original.
B. Result of the median filter.
C. Result of histogram equalization.
Fig. 4Comparison of image resizing methods.
A. Interpolation-based resize (ignore aspect ratio).
B. Zero padding-based resize (maintain aspect ratio).
Fig. 5K-fold cross-validation.
Fig. 6Model compression method for lightweight deep learning models.
A. Weight pruning.
B. Quantization.