Approximating Intermediate Feature Maps of Self-Supervised Convolution Neural Network to Learn Hard Positive Representations in Chest Radiography.

Recent advances in contrastive learning have significantly improved the performance of deep learning models. In contrastive learning of medical images, dealing with positive representation is sometimes difficult because some strong augmentation techniques can disrupt contrastive learning owing to the subtle differences between other standardized CXRs compared to augmented positive pairs; therefore, additional efforts are required. In this study, we propose intermediate feature approximation (IFA) loss, which improves the performance of contrastive convolutional neural networks by focusing more on positive representations of CXRs without additional augmentations. The IFA loss encourages the feature maps of a query image and its positive pair to resemble each other by maximizing the cosine similarity between the intermediate feature outputs of the original data and the positive pairs. Therefore, we used the InfoNCE loss, which is commonly used loss to address negative representations, and the IFA loss, which addresses positive representations, together to improve the contrastive network. We evaluated the performance of the network using various downstream tasks, including classification, object detection, and a generative adversarial network (GAN) inversion task. The downstream task results demonstrated that IFA loss can improve the performance of effectively overcoming data imbalance and data scarcity; furthermore, it can serve as a perceptual loss encoder for GAN inversion. In addition, we have made our model publicly available to facilitate access and encourage further research and collaboration in the field.

Generative AI in glioma: ensuring diversity in training image phenotypes to improve diagnostic performance for IDH mutation prediction.

BACKGROUND: This study evaluated whether generative artificial intelligence-based augmentation (GAA) can provide diverse and realistic imaging phenotypes and improve deep learning-based classification of isocitrate dehydrogenase (IDH) type in glioma compared with neuroradiologists. METHODS: For model development, 565 patients (346 IDH-wildtype, 219 IDH-mutant) with paired contrast-enhanced T1 and FLAIR MRI scans were collected from tertiary hospital and The Cancer Imaging Archive. Performance was tested on internal (119, 78 IDH-wildtype, 41 IDH-mutant [IDH1 and 2]) and external test sets (108, 72 IDH-wildtype, 36 IDH-mutant). GAA was developed using score-based diffusion model and ResNet50 classifier. The optimal GAA was selected in comparison with null model. Two neuroradiologists (R1, R2) assessed realism, diversity of imaging phenotypes, and predicted IDH mutation. The performance of a classifier trained with optimal GAA was compared with that of neuroradiologists using area under the receiver operating characteristics curve (AUC). The effect of tumor size and contrast enhancement on GAA performance was tested. RESULTS: Generated images demonstrated realism (Turing's test: 47.5%-50.5%) and diversity indicating IDH type. Optimal GAA was achieved with augmentation with 110 000 generated slices (AUC: 0.938). The classifier trained with optimal GAA demonstrated significantly higher AUC values than neuroradiologists in both the internal (R1, P=.003; R2, P<.001) and external test sets (R1, P<.01; R2, P<.001). GAA with large-sized tumors or predominant enhancement showed comparable performance to optimal GAA (internal test: AUC 0.956 and 0.922; external test: 0.810 and 0.749). CONCLUSIONS: Application of generative AI with realistic and diverse images provided better diagnostic performance than neuroradiologists for predicting IDH type in glioma.

Orthognathic surgical planning using graph CNN with dual embedding module: External validations with multi-hospital datasets.

BACKGROUND AND OBJECTIVE: Despite recent development of AI, prediction of the surgical movement in the maxilla and mandible by OGS might be more difficult than that of tooth movement by orthodontic treatment. To evaluate the prediction accuracy of the surgical movement using pairs of pre-(T0) and post-surgical (T1) lateral cephalograms (lat-ceph) of orthognathic surgery (OGS) patients and dual embedding module-graph convolution neural network (DEM-GCNN) model. METHODS: 599 pairs from 3 institutions were used as training, internal validation, and internal test sets and 201 pairs from other 6 institutions were used as external test set. DEM-GCNN model (IEM, learning the lat-ceph images; LTEM, learning the landmarks) was developed to predict the amount and direction of surgical movement of ANS and PNS in the maxilla and B-point and Md1crown in the mandible. The distance between T1 landmark coordinates actually moved by OGS (ground truth) and predicted by DEM-GCNN model and pre-existed CNN-based Model-C (learning the lat-ceph images) was compared. RESULTS: In both internal and external tests, DEM-GCNN did not exhibit significant difference from ground truth in all landmarks (ANS, PNS, B-point, Md1crown, all P > 0.05). When the accumulated successful detection rate for each landmark was compared, DEM-GCNN showed higher values than Model-C in both the internal and external tests. In violin plots exhibiting the error distribution of the prediction results, both internal and external tests showed that DEM-GCNN had significant performance improvement in PNS, ANS, B-point, Md1crown than Model-C. DEM-GCNN showed significantly lower prediction error values than Model-C (one-jaw surgery, B-point, Md1crown, all P < 0.005; two-jaw surgery, PNS, ANS, all P < 0.05; B point, Md1crown, all P < 0.005). CONCLUSION: We developed a robust OGS planning model with maximized generalizability despite diverse qualities of lat-cephs from 9 institutions.

Screening obstructive sleep apnea patients via deep learning of knowledge distillation in the lateral cephalogram.

The lateral cephalogram in orthodontics is a valuable screening tool on undetected obstructive sleep apnea (OSA), which can lead to consequences of severe systematic disease. We hypothesized that a deep learning-based classifier might be able to differentiate OSA as anatomical features in lateral cephalogram. Moreover, since the imaging devices used by each hospital could be different, there is a need to overcome modality difference of radiography. Therefore, we proposed a deep learning model with knowledge distillation to classify patients into OSA and non-OSA groups using the lateral cephalogram and to overcome modality differences simultaneously. Lateral cephalograms of 500 OSA patients and 498 non-OSA patients from two different devices were included. ResNet-50 and ResNet-50 with a feature-based knowledge distillation models were trained and their performances of classification were compared. Through the knowledge distillation, area under receiver operating characteristic curve analysis and gradient-weighted class activation mapping of knowledge distillation model exhibits high performance without being deceived by features caused by modality differences. By checking the probability values predicting OSA, an improvement in overcoming the modality differences was observed, which could be applied in the actual clinical situation.

Overcoming the Challenges in the Development and Implementation of Artificial Intelligence in Radiology: A Comprehensive Review of Solutions Beyond Supervised Learning.

Artificial intelligence (AI) in radiology is a rapidly developing field with several prospective clinical studies demonstrating its benefits in clinical practice. In 2022, the Korean Society of Radiology held a forum to discuss the challenges and drawbacks in AI development and implementation. Various barriers hinder the successful application and widespread adoption of AI in radiology, such as limited annotated data, data privacy and security, data heterogeneity, imbalanced data, model interpretability, overfitting, and integration with clinical workflows. In this review, some of the various possible solutions to these challenges are presented and discussed; these include training with longitudinal and multimodal datasets, dense training with multitask learning and multimodal learning, self-supervised contrastive learning, various image modifications and syntheses using generative models, explainable AI, causal learning, federated learning with large data models, and digital twins.

CheSS: Chest X-Ray Pre-trained Model via Self-supervised Contrastive Learning.

Training deep learning models on medical images heavily depends on experts' expensive and laborious manual labels. In addition, these images, labels, and even models themselves are not widely publicly accessible and suffer from various kinds of bias and imbalances. In this paper, chest X-ray pre-trained model via self-supervised contrastive learning (CheSS) was proposed to learn models with various representations in chest radiographs (CXRs). Our contribution is a publicly accessible pretrained model trained with a 4.8-M CXR dataset using self-supervised learning with a contrastive learning and its validation with various kinds of downstream tasks including classification on the 6-class diseases in internal dataset, diseases classification in CheXpert, bone suppression, and nodule generation. When compared to a scratch model, on the 6-class classification test dataset, we achieved 28.5% increase in accuracy. On the CheXpert dataset, we achieved 1.3% increase in mean area under the receiver operating characteristic curve on the full dataset and 11.4% increase only using 1% data in stress test manner. On bone suppression with perceptual loss, we achieved improvement in peak signal to noise ratio from 34.99 to 37.77, structural similarity index measure from 0.976 to 0.977, and root-square-mean error from 4.410 to 3.301 when compared to ImageNet pretrained model. Finally, on nodule generation, we achieved improvement in Fréchet inception distance from 24.06 to 17.07. Our study showed the decent transferability of CheSS weights. CheSS weights can help researchers overcome data imbalance, data shortage, and inaccessibility of medical image datasets. CheSS weight is available at https://github.com/mi2rl/CheSS .

Stain normalization using score-based diffusion model through stain separation and overlapped moving window patch strategies.

Hematoxylin and eosin (H&E) staining is the gold standard modality for diagnosis in medicine. However, the dosage ratio of hematoxylin to eosin in H&E staining has not been standardized yet. Additionally, H&E stains fade out at various speeds. Therefore, the staining quality could differ among each image, and stain normalization is a critical preprocessing approach for training deep learning (DL) models, especially in long-term and/or multicenter digital pathology studies. However, conventional methods for stain normalization have some significant drawbacks, such as collapsing in the structure and/or texture of tissue. In addition, conventional methods must require a reference patch or slide. Meanwhile, DL-based methods have a risk of overfitting and/or grid artifacts. We developed a score-based diffusion model of colorization for stain normalization. However, mistransfer, in which the model confuses hematoxylin with eosin, can occur using a score-based diffusion model due to its high diversity nature. To overcome this mistransfer, we propose a stain separation method using sparse non-negative matrix factorization (SNMF), which can decompose pathology slide into Hematoxylin and Eosin to normalize each stain component. Furthermore, inpainting with overlapped moving window patches was used to prevent grid artifacts of whole slide image normalization. Our method can normalize the whole slide pathology images through this stain normalization pipeline with decent performance.

Tat-PRAS40 prevent hippocampal HT-22 cell death and oxidative stress induced animal brain ischemic insults.

Proline rich Akt substrate (PRAS40) is a component of mammalian target of rapamycin complex 1 (mTORC1) and is known to play an important role against reactive oxygen species-induced cell death. However, the precise function of PRAS40 in ischemia remains unclear. Thus, we investigated whether Tat-PRAS40, a cell-permeable fusion protein, has a protective function against oxidative stress-induced hippocampal neuronal (HT-22) cell death in an animal model of ischemia. We showed that Tat-PRAS40 transduced into HT-22 cells, and significantly protected against cell death by reducing the levels of H2O2 and derived reactive species, and DNA fragmentation as well as via the regulation of Bcl-2, Bax, and caspase 3 expression levels in H2O2 treated cells. Also, we showed that transduced Tat-PARS40 protein markedly increased phosphorylated RRAS40 expression levels and 14-3-3σ complex via the Akt signaling pathway. In an animal ischemia model, Tat-PRAS40 effectively transduced into the hippocampus in animal brain and significantly protected against neuronal cell death in the CA1 region. We showed that Tat-PRAS40 protein effectively transduced into hippocampal neuronal cells and markedly protected against neuronal cell damage. Therefore, we suggest that Tat-PRAS40 protein may be used as a therapeutic protein for ischemia and oxidative stress-induced brain disorders.

Tat-CBR1 inhibits inflammatory responses through the suppressions of NF-κB and MAPK activation in macrophages and TPA-induced ear edema in mice.

Human carbonyl reductase 1 (CBR1) plays a crucial role in cell survival and protects against oxidative stress response. However, its anti-inflammatory effects are not yet clearly understood. In this study, we examined whether CBR1 protects against inflammatory responses in macrophages and mice using a Tat-CBR1 protein which is able to penetrate into cells. The results revealed that purified Tat-CBR1 protein efficiently transduced into Raw 264.7 cells and inhibited lipopolysaccharide (LPS)-induced cyclooxygenase-2 (COX-2), nitric oxide (NO) and prostaglandin E2 (PGE2) expression levels. In addition, Tat-CBR1 protein leads to decreased pro-inflammatory cytokine expression through suppression of nuclear transcription factor-kappaB (NF-κB) and mitogen activated protein kinase (MAPK) activation. Furthermore, Tat-CBR1 protein inhibited inflammatory responses in 12-O-tetradecanoylphorbol-13-acetate (TPA)-induced skin inflammation when applied topically. These findings indicate that Tat-CBR1 protein has anti-inflammatory properties in vitro and in vivo through inhibition of NF-κB and MAPK activation, suggesting that Tat-CBR1 protein may have potential as a therapeutic agent against inflammatory diseases.

Down-regulation of MAPK/NF-κB signaling underlies anti-inflammatory response induced by transduced PEP-1-Prx2 proteins in LPS-induced Raw 264.7 and TPA-induced mouse ear edema model.

Excessive reactive oxygen species (ROS) production plays a crucial role in causing various diseases, including inflammatory disorders. The activation of mitogen-activated protein kinase (MAPK) and nuclear factor-kappaB (NF-κB) signaling is implicated in stimulating inflammatory response and cytokines. Peroxiredoxin 2 (Prx2) is a 2-cysteine (Cys) peroxiredoxin capable of removing endogenous hydrogen peroxide (H2O2). PEP-1 peptide, a protein transduction domain, consists of three domains which are used to transduce exogenous therapeutic proteins into cells. The correlation between effectively transduced PEP-1-Prx2 and ROS-mediated inflammatory response is not clear. In the present study, we investigated the protective effects of cell permeable PEP-1-Prx2 on oxidative stress-induced inflammatory activity in Raw 264.7 cells and in a mouse ear edema model after exposure to lipopolysaccharides (LPS) or 12-O-tetradecanoylphorbol-13-acetate (TPA). Transduced PEP-1-Prx2 suppressed intracellular ROS accumulation and inhibited the activity of MAPKs and NF-κB signaling that led to the suppression of cyclooxygenase-2 (COX-2), inducible nitric oxide synthase (iNOS) and cytokines in LPS-induced Raw 264.7 cells and TPA-induced mouse ear edema model. Given these results, we propose that PEP-1-Prx2 has therapeutic potential in the prevention of inflammatory disorders.

Altered expression of adrenocorticotropic hormone in the epileptic gerbil hippocampus following spontaneous seizure.

We investigated the temporal alterations of adrenocorticotropic hormone (ACTH) immunoreactivity in the hippocampus after seizure onset. Expression of ACTH was observed within interneurons in the pre-seizure group of seizure sensitive gerbils, whereas its immunoreactivities were rarely detected in seizure resistant gerbil. Three hr after the seizure, ACTH immunoreactivity was significantly increased in interneurons within all hippocampal regions. On the basis of their localization and morphology through immunofluorescence staining, these cells were identified as GABAA α1-containing interneurons. At the 12 hr postictal period, ACTH expression in these regions was down-regulated, in a similar manner to the pre-seizure group of gerbils. These findings support the increase in ACTH synthesis that contributes to a reduction of corticotrophin-releasing factor via the negative feedback system which in turn provides an opportunity to enhance the excitability of GABAergic interneurons. Therefore, ACTH may play an important role in the reduction of excitotoxicity in all hippocampal regions.

Alterations in hyperpolarization-activated cyclic nucleotidegated cation channel (HCN) expression in the hippocampus following pilocarpine-induced status epilepticus.

To understand the effects of HCN as potential mediators in the pathogenesis of epilepsy that evoke long-term impaired excitability; the present study was designed to elucidate whether the alterations of HCN expression induced by status epilepticus (SE) is responsible for epileptogenesis. Although HCN1 immunoreactivity was observed in the hippocampus, its immunoreactivities were enhanced at 12 hrs following SE. Although, HCN1 immunoreactivities were reduced in all the hippocampi at 2 weeks, a re-increase in the expression at 2-3 months following SE was observed. In contrast to HCN1, HCN 4 expressions were un-changed, although HCN2 immunoreactive neurons exhibited some changes following SE. Taken together, our findings suggest that altered expressions of HCN1 following SE may be mainly involved in the imbalances of neurotransmissions to hippocampal circuits; thus, it is proposed that HCN1 may play an important role in the epileptogenic period as a compensatory response.

Altered PLCß-1 expression in the gerbil hippocampal complex following spontaneous seizure.

Although the phospholipase C (PLC)ß-1 isoform is associated with spontaneous seizure and distinctively expressed in the telencephalon, the distribution of PLCß-1 expression in the epileptic gerbil hippocampus remains controversial. Therefore, we determined whether PLCß-1 is associated with spontaneous seizure in an animal model of genetic epilepsy. In the present study, PLCß-1 immunoreactivity was down-regulated in seizure-sensitive (SS) gerbils more than in seizure-resistant (SR) gerbils. The expression of PLCß-1 within calretinin (CR)- positive neurons was rarely detected within the dentate hilar region of SS gerbils. PLCß-1 immunoreactivity in the hippocampus was significantly elevated as compared to that in pre-seizure SS gerbil 3 h post-ictal. These findings suggest that alterations in PLCß-1 immunoreactivity in the SS gerbil hippocampus may be closely related to the epileptic state of the gerbil brain and transiently elevated PLCß-1 protein levels following seizure episodes. Such alterations may be compensatory responses in the SS gerbil hippocampus.