Impact of color augmentation and tissue type in deep learning for hematoxylin and eosin image super resolution.

Single image super-resolution is an important computer vision task with applications including remote sensing, medical imaging, and surveillance. Modern work on super-resolution utilizes deep learning to synthesize high resolution (HR) images from low resolution images (LR). With the increased utilization of digitized whole slide images (WSI) in pathology workflows, digital pathology has emerged as a promising domain for super-resolution. Despite extensive existing research into super-resolution, there remain challenges specific to digital pathology. Here, we investigated image augmentation techniques for hematoxylin and eosin (H&E) WSI super-resolution and model generalizability across diverse tissue types. In addition, we investigated shortcomings with common quality metrics (peak signal-to-noise ratio (PSNR), structure similarity index (SSIM)) by conducting a perceptual quality survey for super-resolved pathology images. High performing deep super-resolution models were used to generate 20X HR images from LR images (5X or 10X equivalent) for 11 different tissues and 30 human evaluators were asked to score the quality of the generated versus the ground truth 20X HR images. The scores given by a human rater and the PSNR or the SSIM were compared to investigate the correlation between model training parameters. We found that models trained on multiple tissues generalized better than those trained on a single tissue type. We also found that PSNR correlated with perceptual quality (R = 0.26) less accurately than did SSIM (R = 0.64), suggesting that the SSIM quality metric is insufficient. The methods proposed in this study can be used to virtually magnify H&E images with better perceptual quality than interpolation methods (i.e., bicubic interpolation) commonly implemented in digital pathology software. The impact of deep SISR methods is more notable when scaling to 4X is needed, such as in the case of super-resolving a low magnification WSI from 10X to 40X.

Multiscale generative model using regularized skip-connections and perceptual loss for anomaly detection in toxicologic histopathology.

Background: Automated anomaly detection is an important tool that has been developed for many real-world applications, including security systems, industrial inspection, and medical diagnostics. Despite extensive use of machine learning for anomaly detection in these varied contexts, it is challenging to generalize and apply these methods to complex tasks such as toxicologic histopathology (TOXPATH) assessment (i.e.,finding abnormalities in organ tissues). In this work, we introduce an anomaly detection method using deep learning that greatly improves model generalizability to TOXPATH data. Methods: We evaluated a one-class classification approach that leverages novel regularization and perceptual techniques within generative adversarial network (GAN) and autoencoder architectures to accurately detect anomalous histopathological findings of varying degrees of complexity. We also utilized multiscale contextual data and conducted a thorough ablation study to demonstrate the efficacy of our method. We trained our models on data from normal whole slide images (WSIs) of rat liver sections and validated on WSIs from three anomalous classes. Anomaly scores are collated into heatmaps to localize anomalies within WSIs and provide human-interpretable results. Results: Our method achieves 0.953 area under the receiver operating characteristic on a real-worldTOXPATH dataset. The model also shows good performance at detecting a wide variety of anomalies demonstrating our method's ability to generalize to TOXPATH data. Conclusion: Anomalies in both TOXPATH histological and non-histological datasets were accurately identified with our method, which was only trained with normal data.

Assessment of Skin Toxicity in an in Vitro Reconstituted Human Epidermis Model Using Deep Learning.

Skin toxicity is a common safety concern associated with drugs that inhibit epidermal growth factor receptors as well as other targets involved in epidermal growth and differentiation. Recently, the use of a three-dimensional reconstructed human epidermis model enabled large-scale drug screening and showed potential for predicting skin toxicity. Although a decrease in epidermal thickness was often observed when the three-dimensional reconstructed tissues were exposed to drugs causing skin toxicity, the thickness evaluation of epidermal layers from a pathologist was subjective and not easily reproducible or scalable. In addition, the subtle differences in thickness among tissues, as well as the large number of samples tested, made cross-study comparison difficult when a manual evaluation strategy was used. The current study used deep learning and image-processing algorithms to measure the viable epidermal thickness from multiple studies and found that the measured thickness was not only significantly correlated with a pathologist's semi-quantitative evaluation but was also in close agreement with the quantitative measurement performed by pathologists. Moreover, a sensitivity of 0.8 and a specificity of 0.75 were achieved when predicting the toxicity of 18 compounds with clinical observations with these epidermal thickness algorithms. This approach is fully automated, reproducible, and highly scalable. It not only shows reasonable accuracy in predicting skin toxicity but also enables cross-study comparison and high-throughput compound screening.