Advancing Medical Imaging Informatics by Deep Learning-Based Domain Adaptation.

INTRODUCTION: There has been a rapid development of deep learning (DL) models for medical imaging. However, DL requires a large labeled dataset for training the models. Getting large-scale labeled data remains a challenge, and multi-center datasets suffer from heterogeneity due to patient diversity and varying imaging protocols. Domain adaptation (DA) has been developed to transfer the knowledge from a labeled data domain to a related but unlabeled domain in either image space or feature space. DA is a type of transfer learning (TL) that can improve the performance of models when applied to multiple different datasets. OBJECTIVE: In this survey, we review the state-of-the-art DL-based DA methods for medical imaging. We aim to summarize recent advances, highlighting the motivation, challenges, and opportunities, and to discuss promising directions for future work in DA for medical imaging. METHODS: We surveyed peer-reviewed publications from leading biomedical journals and conferences between 2017-2020, that reported the use of DA in medical imaging applications, grouping them by methodology, image modality, and learning scenarios. RESULTS: We mainly focused on pathology and radiology as application areas. Among various DA approaches, we discussed domain transformation (DT) and latent feature-space transformation (LFST). We highlighted the role of unsupervised DA in image segmentation and described opportunities for future development. CONCLUSION: DA has emerged as a promising solution to deal with the lack of annotated training data. Using adversarial techniques, unsupervised DA has achieved good performance, especially for segmentation tasks. Opportunities include domain transferability, multi-modal DA, and applications that benefit from synthetic data.

Learning to Evaluate Color Similarity for Histopathology Images using Triplet Networks.

Stain normalization is a crucial pre-processing step for histopathological image processing, and can help improve the accuracy of downstream tasks such as segmentation and classification. To evaluate the effectiveness of stain normalization methods, various metrics based on color-perceptual similarity and stain color evaluation have been proposed. However, there still exists a huge gap between metric evaluation and human perception, given the limited explainability power of existing metrics and inability to combine color and semantic information efficiently. Inspired by the effectiveness of deep neural networks in evaluating perceptual similarity of natural images, in this paper, we propose TriNet-P, a color-perceptual similarity metric for whole slide images, based on deep metric embeddings. We evaluate the proposed approach using four publicly available breast cancer histological datasets. The benefit of our approach is its representation efficiency of the perceptual factors associated with H&E stained images with minimal human intervention. We show that our metric can capture the semantic similarities, both at subject (patient) and laboratory levels, and leads to better performance in image retrieval and clustering tasks.

Textural characterization of histopathological images for oral sub-mucous fibrosis detection.

In the field of quantitative microscopy, textural information plays a significant role very often in tissue characterization and diagnosis, in addition to morphology and intensity. The aim of this work is to improve the classification accuracy based on textural features for the development of a computer assisted screening of oral sub-mucous fibrosis (OSF). In fact, a systematic approach is introduced in order to grade the histopathological tissue sections into normal, OSF without dysplasia and OSF with dysplasia, which would help the oral onco-pathologists to screen the subjects rapidly. In totality, 71 textural features are extracted from epithelial region of the tissue sections using various wavelet families, Gabor-wavelet, local binary pattern, fractal dimension and Brownian motion curve, followed by preprocessing and segmentation. Wavelet families contribute a common set of 9 features, out of which 8 are significant and other 61 out of 62 obtained from the rest of the extractors are also statistically significant (p<0.05) in discriminating the three stages. Based on mean distance criteria, the best wavelet family (i.e., biorthogonal3.1 (bior3.1)) is selected for classifier design. support vector machine (SVM) is trained by 146 samples based on 69 textural features and its classification accuracy is computed for each of the combinations of wavelet family and rest of the extractors. Finally, it has been investigated that bior3.1 wavelet coefficients leads to higher accuracy (88.38%) in combination with LBP and Gabor wavelet features through three-fold cross validation. Results are shown and discussed in detail. It is shown that combining more than one texture measure instead of using just one might improve the overall accuracy.

Texture based segmentation of epithelial layer from oral histological images.

The objective of this paper is to provide a texture based segmentation algorithm for better delineation of the epithelial layer from histological images in discriminating normal and oral sub-mucous fibrosis (OSF). As per literature and oral clinicians, it is established that the OSF initially originates and propagates in the epithelial layer. So, more accurate segmentation of this layer is extremely important for a clinician to make a diagnostic decision. In doing this, Gabor based texture gradient is computed in gray scale images, followed by preprocessing of the microscopic images of oral histological sections. On the other hand, the color gradients of these images are obtained in the transformed Lab color space. Finally, the watershed segmentation is extended to segment the layer based on the combination of texture and color gradients. The segmented images are compared with the ground truth images provided by the oral experts. The segmentation results depict the superiority of the texture based segmentation in comparison to the Otsu's based segmentation in terms of misclassification error. Results are shown and discussed.