Exploration of Interpretability Techniques for Deep COVID-19 Classification Using Chest X-ray Images

The outbreak of COVID-19 has shocked the entire world with its fairly rapid spread, and has challenged different sectors. One of the most effective ways to limit its spread is the early and accurate diagnosing of infected patients. Medical imaging, such as X-ray and computed tomography (CT), combined with the potential of artificial intelligence (AI), plays an essential role in supporting medical personnel in the diagnosis process. Thus, in this article, five different deep learning models (ResNet18, ResNet34, InceptionV3, InceptionResNetV2, and DenseNet161) and their ensemble, using majority voting, have been used to classify COVID-19, pneumoniæ and healthy subjects using chest X-ray images. Multilabel classification was performed to predict multiple pathologies for each patient, if present. Firstly, the interpretability of each of the networks was thoroughly studied using local interpretability methods-occlusion, saliency, input X gradient, guided backpropagation, integrated gradients, and DeepLIFT-and using a global technique-neuron activation profiles. The mean micro F1 score of the models for COVID-19 classifications ranged from 0.66 to 0.875, and was 0.89 for the ensemble of the network models. The qualitative results showed that the ResNets were the most interpretable models. This research demonstrates the importance of using interpretability methods to compare different models before making a decision regarding the best performing model.

An Ensemble of Deep Learning Object Detection Models for Anatomical and Pathological Regions in Brain MRI

This paper proposes ensemble strategies for the deep learning object detection models carried out by combining the variants of a model and different models to enhance the anatomical and pathological object detection performance in brain MRI. In this study, with the help of the novel Gazi Brains 2020 dataset, five different anatomical parts and one pathological part that can be observed in brain MRI were identified, such as the region of interest, eye, optic nerves, lateral ventricles, third ventricle, and a whole tumor. Firstly, comprehensive benchmarking of the nine state-of-the-art object detection models was carried out to determine the capabilities of the models in detecting the anatomical and pathological parts. Then, four different ensemble strategies for nine object detectors were applied to boost the detection performance using the bounding box fusion technique. The ensemble of individual model variants increased the anatomical and pathological object detection performance by up to 10% in terms of the mean average precision (mAP). In addition, considering the class-based average precision (AP) value of the anatomical parts, an up to 18% AP improvement was achieved. Similarly, the ensemble strategy of the best different models outperformed the best individual model by 3.3% mAP. Additionally, while an up to 7% better FAUC, which is the area under the TPR vs. FPPI curve, was achieved on the Gazi Brains 2020 dataset, a 2% better FAUC score was obtained on the BraTS 2020 dataset. The proposed ensemble strategies were found to be much more efficient in finding the anatomical and pathological parts with a small number of anatomic objects, such as the optic nerve and third ventricle, and producing higher TPR values, especially at low FPPI values, compared to the best individual methods.

Visual Diagnostics of Dental Caries through Deep Learning of Non-Standardised Photographs Using a Hybrid YOLO Ensemble and Transfer Learning Model

BACKGROUND

Access to oral healthcare is not uniform globally, particularly in rural areas with limited resources, which limits the potential of automated diagnostics and advanced tele-dentistry applications. The use of digital caries detection and progression monitoring through photographic communication, is influenced by multiple variables that are difficult to standardize in such settings. The objective of this study was to develop a novel and cost-effective virtual computer vision AI system to predict dental cavitations from non-standardised photographs with reasonable clinical accuracy.

METHODS

A set of 1703 augmented images was obtained from 233 de-identified teeth specimens. Images were acquired using a consumer smartphone, without any standardised apparatus applied. The study utilised state-of-the-art ensemble modeling, test-time augmentation, and transfer learning processes. The "you only look once" algorithm (YOLO) derivatives, v5s, v5m, v5l, and v5x, were independently evaluated, and an ensemble of the best results was augmented, and transfer learned with ResNet50, ResNet101, VGG16, AlexNet, and DenseNet. The outcomes were evaluated using precision, recall, and mean average precision (mAP).

RESULTS

The YOLO model ensemble achieved a mean average precision (mAP) of 0.732, an accuracy of 0.789, and a recall of 0.701. When transferred to VGG16, the final model demonstrated a diagnostic accuracy of 86.96%, precision of 0.89, and recall of 0.88. This surpassed all other base methods of object detection from free-hand non-standardised smartphone photographs.

CONCLUSION

A virtual computer vision AI system, blending a model ensemble, test-time augmentation, and transferred deep learning processes, was developed to predict dental cavitations from non-standardised photographs with reasonable clinical accuracy. This model can improve access to oral healthcare in rural areas with limited resources, and has the potential to aid in automated diagnostics and advanced tele-dentistry applications.

MuLan-Methyl-multiple transformer-based language models for accurate DNA methylation prediction

Transformer-based language models are successfully used to address massive text-related tasks. DNA methylation is an important epigenetic mechanism, and its analysis provides valuable insights into gene regulation and biomarker identification. Several deep learning-based methods have been proposed to identify DNA methylation, and each seeks to strike a balance between computational effort and accuracy. Here, we introduce MuLan-Methyl, a deep learning framework for predicting DNA methylation sites, which is based on 5 popular transformer-based language models. The framework identifies methylation sites for 3 different types of DNA methylation: N6-adenine, N4-cytosine, and 5-hydroxymethylcytosine. Each of the employed language models is adapted to the task using the "pretrain and fine-tune" paradigm. Pretraining is performed on a custom corpus of DNA fragments and taxonomy lineages using self-supervised learning. Fine-tuning aims at predicting the DNA methylation status of each type. The 5 models are used to collectively predict the DNA methylation status. We report excellent performance of MuLan-Methyl on a benchmark dataset. Moreover, we argue that the model captures characteristic differences between different species that are relevant for methylation. This work demonstrates that language models can be successfully adapted to applications in biological sequence analysis and that joint utilization of different language models improves model performance. Mulan-Methyl is open source, and we provide a web server that implements the approach.

Accuracy, realism and general applicability of European forest models

Forest models are instrumental for understanding and projecting the impact of climate change on forests. A considerable number of forest models have been developed in the last decades. However, few systematic and comprehensive model comparisons have been performed in Europe that combine an evaluation of modelled carbon and water fluxes and forest structure. We evaluate 13 widely used, state-of-the-art, stand-scale forest models against field measurements of forest structure and eddy-covariance data of carbon and water fluxes over multiple decades across an environmental gradient at nine typical European forest stands. We test the models' performance in three dimensions: accuracy of local predictions (agreement of modelled and observed annual data), realism of environmental responses (agreement of modelled and observed responses of daily gross primary productivity to temperature, radiation and vapour pressure deficit) and general applicability (proportion of European tree species covered). We find that multiple models are available that excel according to our three dimensions of model performance. For the accuracy of local predictions, variables related to forest structure have lower random and systematic errors than annual carbon and water flux variables. Moreover, the multi-model ensemble mean provided overall more realistic daily productivity responses to environmental drivers across all sites than any single individual model. The general applicability of the models is high, as almost all models are currently able to cover Europe's common tree species. We show that forest models complement each other in their response to environmental drivers and that there are several cases in which individual models outperform the model ensemble. Our framework provides a first step to capturing essential differences between forest models that go beyond the most commonly used accuracy of predictions. Overall, this study provides a point of reference for future model work aimed at predicting climate impacts and supporting climate mitigation and adaptation measures in forests.

NonClasGP-Pred: robust and efficient prediction of non-classically secreted proteins by integrating subset-specific optimal models of imbalanced data

Non-classically secreted proteins (NCSPs) are proteins that are located in the extracellular environment, although there is a lack of known signal peptides or secretion motifs. They usually perform different biological functions in intracellular and extracellular environments, and several of their biological functions are linked to bacterial virulence and cell defence. Accurate protein localization is essential for all living organisms, however, the performance of existing methods developed for NCSP identification has been unsatisfactory and in particular suffer from data deficiency and possible overfitting problems. Further improvement is desirable, especially to address the lack of informative features and mining subset-specific features in imbalanced datasets. In the present study, a new computational predictor was developed for NCSP prediction of gram-positive bacteria. First, to address the possible prediction bias caused by the data imbalance problem, ten balanced subdatasets were generated for ensemble model construction. Then, the F-score algorithm combined with sequential forward search was used to strengthen the feature representation ability for each of the training subdatasets. Third, the subset-specific optimal feature combination process was adopted to characterize the original data from different aspects, and all subdataset-based models were integrated into a unified model, NonClasGP-Pred, which achieved an excellent performance with an accuracy of 93.23 %, a sensitivity of 100 %, a specificity of 89.01 %, a Matthew's correlation coefficient of 87.68 % and an area under the curve value of 0.9975 for ten-fold cross-validation. Based on assessment on the independent test dataset, the proposed model outperformed state-of-the-art available toolkits. For availability and implementation, see: http://lab.malab.cn/~wangchao/softwares/NonClasGP/.

Climate change promotes transitions to tall evergreen vegetation in tropical Asia

Vegetation in tropical Asia is highly diverse due to large environmental gradients and heterogeneity of landscapes. This biodiversity is threatened by intense land use and climate change. However, despite the rich biodiversity and the dense human population, tropical Asia is often underrepresented in global biodiversity assessments. Understanding how climate change influences the remaining areas of natural vegetation is therefore highly important for conservation planning. Here, we used the adaptive Dynamic Global Vegetation Model version 2 (aDGVM2) to simulate impacts of climate change and elevated CO₂ on vegetation formations in tropical Asia for an ensemble of climate change scenarios. We used climate forcing from five different climate models for representative concentration pathways RCP4.5 and RCP8.5. We found that vegetation in tropical Asia will remain a carbon sink until 2099, and that vegetation biomass increases of up to 28% by 2099 are associated with transitions from small to tall woody vegetation and from deciduous to evergreen vegetation. Patterns of phenology were less responsive to climate change and elevated CO₂ than biomes and biomass, indicating that the selection of variables and methods used to detect vegetation changes is crucial. Model simulations revealed substantial variation within the ensemble, both in biomass increases and in distributions of different biome types. Our results have important implications for management policy, because they suggest that large ensembles of climate models and scenarios are required to assess a wide range of potential future trajectories of vegetation change and to develop robust management plans. Furthermore, our results highlight open ecosystems with low tree cover as most threatened by climate change, indicating potential conflicts of interest between biodiversity conservation in open ecosystems and active afforestation to enhance carbon sequestration.

Global wheat production with 1.5 and 2.0°C above pre-industrial warming

Efforts to limit global warming to below 2°C in relation to the pre-industrial level are under way, in accordance with the 2015 Paris Agreement. However, most impact research on agriculture to date has focused on impacts of warming >2°C on mean crop yields, and many previous studies did not focus sufficiently on extreme events and yield interannual variability. Here, with the latest climate scenarios from the Half a degree Additional warming, Prognosis and Projected Impacts (HAPPI) project, we evaluated the impacts of the 2015 Paris Agreement range of global warming (1.5 and 2.0°C warming above the pre-industrial period) on global wheat production and local yield variability. A multi-crop and multi-climate model ensemble over a global network of sites developed by the Agricultural Model Intercomparison and Improvement Project (AgMIP) for Wheat was used to represent major rainfed and irrigated wheat cropping systems. Results show that projected global wheat production will change by -2.3% to 7.0% under the 1.5°C scenario and -2.4% to 10.5% under the 2.0°C scenario, compared to a baseline of 1980-2010, when considering changes in local temperature, rainfall, and global atmospheric CO₂ concentration, but no changes in management or wheat cultivars. The projected impact on wheat production varies spatially; a larger increase is projected for temperate high rainfall regions than for moderate hot low rainfall and irrigated regions. Grain yields in warmer regions are more likely to be reduced than in cooler regions. Despite mostly positive impacts on global average grain yields, the frequency of extremely low yields (bottom 5 percentile of baseline distribution) and yield inter-annual variability will increase under both warming scenarios for some of the hot growing locations, including locations from the second largest global wheat producer-India, which supplies more than 14% of global wheat. The projected global impact of warming <2°C on wheat production is therefore not evenly distributed and will affect regional food security across the globe as well as food prices and trade.