Recent developments in deep learning have impacted medical science. However, new privacy issues and regulatory frameworks have hindered medical data sharing and collection. Deep learning is a very data-intensive process for which such regulatory limitations limit the potential for new breakthroughs and collaborations. However, generating medically accurate synthetic data can alleviate privacy issues and potentially augment deep learning pipelines. This study presents generative adversarial neural networks capable of generating realistic images of knee joint X-rays with varying osteoarthritis severity. We offer 320,000 synthetic (DeepFake) X-ray images from training with 5,556 real images. We validated our models regarding medical accuracy with 15 medical experts and for augmentation effects with an osteoarthritis severity classification task. We devised a survey of 30 real and 30 DeepFake images for medical experts. The result showed that on average, more DeepFakes were mistaken for real than the reverse. The result signified sufficient DeepFake realism for deceiving the medical experts. Finally, our DeepFakes improved classification accuracy in an osteoarthritis severity classification task with scarce real data and transfer learning. In addition, in the same classification task, we replaced all real training data with DeepFakes and suffered only a $$3.79\%$$3.79% loss from baseline accuracy in classifying real osteoarthritis X-rays.

Hematoxylin and eosin-stained biopsy slides are regularly available for colorectal cancer patients. These slides are often not used to define objective biomarkers for patient stratification and treatment selection. Standard biomarkers often pertain to costly and slow genetic tests. However, recent work has shown that relevant biomarkers can be extracted from these images using convolutional neural networks (CNNs). The CNN-based biomarkers predicted colorectal cancer patient outcomes comparably to gold standards. Extracting CNN-biomarkers is fast, automatic, and of minimal cost. CNN-based biomarkers rely on the ability of CNNs to recognize distinct tissue types from microscope whole slide images. The quality of these biomarkers (coined ‘Deep Stroma’) depends on the accuracy of CNNs in decomposing all relevant tissue classes. Improving tissue decomposition accuracy is essential for improving the prognostic potential of CNN-biomarkers. In this study, we implemented a novel training strategy to refine an established CNN model, which then surpassed all previous solutions . We obtained a 95.6% average accuracy in the external test set and 99.5% in the internal test set. Our approach reduced errors in biomarker-relevant classes, such as Lymphocytes, and was the first to include interpretability methods. These methods were used to better apprehend our model’s limitations and capabilities.

In diagnostic histopathology, hematoxylin and eosin (H&E) staining is a critical process that highlights salient histological features. Staining results vary between laboratories regardless of the histopathological task, although the method does not change. This variance can impair the accuracy of algorithms and histopathologists’ time-to-insight. Investigating this variance can help calibrate stain normalization tasks to reverse this negative potential. With machine learning, this study evaluated the staining variance between different laboratories on three tissue types. We received H&E-stained slides from 66 different laboratories. Each slide contained kidney, skin, and colon tissue samples stained by the method routinely used in each laboratory. The samples were digitized and summarized as red, green, and blue channel histograms. Dimensions were reduced using principal component analysis. The data projected by principal components were inserted into the k-means clustering algorithm and the k-nearest neighbors classifier with the laboratories as the target. The k-means silhouette index indicated that K = 2 clusters had the best separability in all tissue types. The supervised classification result showed laboratory effects and tissue-type bias. Both supervised and unsupervised approaches suggested that tissue type also affected inter-laboratory variance. We suggest tissue type to also be considered upon choosing the staining and color-normalization approach.

Diagnosing knee joint osteoarthritis (KOA), a major cause of disability worldwide, is challenging due to subtle radiographic indicators and the varied progression of the disease. Using deep learning for KOA diagnosis requires broad, comprehensive datasets. However, obtaining these datasets poses significant challenges due to patient privacy and data collection restrictions. Additive data augmentation, which enhances data variability, emerges as a promising solution. Yet, it’s unclear which augmentation techniques are most effective for KOA. Our study explored data augmentation methods, including adversarial techniques. We used strategies like horizontal cropping and region of interest (ROI) extraction, alongside adversarial methods such as noise injection and ROI removal. Interestingly, rotations improved performance, while methods like horizontal split were less effective. We discovered potential confounding regions using adversarial augmentation, shown in our models’ accurate classification of extreme KOA grades, even without the knee joint. This indicated a potential model bias towards irrelevant radiographic features. Removing the knee joint paradoxically increased accuracy in classifying early-stage KOA. Grad-CAM visualizations helped elucidate these effects. Our study contributed to the field by pinpointing augmentation techniques that either improve or impede model performance, in addition to recognizing potential confounding regions within radiographic images of knee osteoarthritis.

Knee Osteoarthritis (KOA), a leading cause of disability worldwide, is challenging to detect early due to subtle radiographic indicators. Diverse, extensive datasets are needed but are challenging to compile because of privacy, data collection limitations, and the progressive nature of KOA. However, a model capable of projecting genuine radiographs into different OA stages could augment data pools, enhance algorithm training, and offer pre-emptive prognostic insights. In this study, we developed a Cycle-Consistent Adversarial Network (CycleGAN) to generate synthetic past and future stages of KOA on any genuine radiograph. The model's effectiveness was validated through its impact on a KOA specialized Convolutional Neural Network (CNN). Transformations towards synthetic future disease states resulted in 83.76% of none-to-doubtful stage images being classified as moderate-to-severe stages, while retroactive transformations led to 75.61% of severe-stage images being classified as none-to-doubtful stages. Similarly, transformations from mild stages achieved 76.00% correct classification towards future stages and 69.00% for past stages. The CycleGAN demonstrated an exceptional ability to expand the knee joint space and eliminate bone-outgrowths (osteophytes), key radiographic indicators of disease progression. These results signify a promising potential for enhancing diagnostic models, data augmentation, and educational and prognostic uses. Nevertheless, further refinement, validation, and a broader evaluation process encompassing both CNN-based assessments and expert medical feedback are emphasized for future research and development.

Diagnosing knee joint osteoarthritis (KOA), a major cause of disability worldwide, is challenging due to subtle radiographic indicators and the varied progression of the disease. Using deep learning for KOA diagnosis requires broad, comprehensive datasets. However, obtaining these datasets poses significant challenges due to patient privacy concerns and data collection restrictions. Additive data augmentation, which enhances data variability, emerges as a promising solution. Yet, it's unclear which augmentation techniques are most effective for KOA. This study explored various data augmentation methods, including adversarial augmentations, and their impact on KOA classification model performance. While some techniques improved performance, others commonly used underperformed. We identified potential confounding regions within the images using adversarial augmentation. This was evidenced by our models' ability to classify KL0 and KL4 grades accurately, with the knee joint omitted. This observation suggested a model bias, which might leverage unrelated features for classification currently present in radiographs. Interestingly, removing the knee joint also led to an unexpected improvement in KL1 classification accuracy. To better visualize these paradoxical effects, we employed Grad-CAM, highlighting the associated regions. Our study underscores the need for careful technique selection for improved model performance and identifying and managing potential confounding regions in radiographic KOA deep learning.