Confidence In Predictions Research Articles

LesionMix data enhancement and entropy minimization for semi-supervised lesion segmentation of lung cancer

Determining the location and contour of the lesion is a crucial prerequisite for medical diagnosis, subsequent personalized treatment plan and prognostic prediction of lung cancer. Semi-supervised learning and data augmentation methods facilitate deep learning to be used in many fields of medical imaging. In this paper, we introduce a novel data enhancement technique called LesionMix. This method involves extracting and reusing lesions from a limited amount of labeled CT data, thereby enhancing the efficiency of utilizing those labeled data. Meanwhile, we propose a two-stage semi-supervised training strategy called Entropy Minimization LesionMix (EMLM). In the first stage, features containing lesion contour information are rapidly learned through LesionMix data augmentation. Entropy minimization strategy optimizes the model parameters to alleviate overfitting as much as possible in the second stage and improves prediction confidence. Our proposed method is validated on the public dataset LIDC-IDRI and the in-house dataset GDPHLUAD. Extensive experiments demonstrate that our method achieves promising performance and outperforms seven state-of-the-art semi-supervised models; Moreover, ablation experiments validate the effectiveness of various aspects of our approach.

SMoRe GloS: An efficient and flexible framework for inferring global sensitivity of agent-based model parameters.

Agent-based models (ABMs) have become essential tools for simulating complex biological, ecological, and social systems where emergent behaviors arise from the interactions among individual agents. Quantifying uncertainty through global sensitivity analysis is crucial for assessing the robustness and reliability of ABM predictions. However, most global sensitivity methods demand substantial computational resources, making them impractical for highly complex models. Here, we introduce SMoRe GloS (Surrogate Modeling for Recapitulating Global Sensitivity), a novel, computationally efficient method for performing global sensitivity analysis of ABMs. By leveraging explicitly formulated surrogate models, SMoRe GloS allows for comprehensive parameter space exploration and uncertainty quantification without sacrificing accuracy. We demonstrate our method's flexibility by applying it to two biological ABMs: a simple 2D cell proliferation assay and a complex 3D vascular tumor growth model. Our results show that SMoRe GloS is compatible with simpler methods like the Morris one-at-a-time method, and more computationally intensive variance-based methods like eFAST. SMoRe GloS accurately recovered global sensitivity indices in each case while achieving substantial speedups, completing analyses in minutes. In contrast, direct implementation of eFAST amounted to several days of CPU time for the complex ABM. Remarkably, our method also estimates sensitivities for ABM parameters representing processes not explicitly included in the surrogate model, further enhancing its utility. By making global sensitivity analysis feasible for computationally expensive models, SMoRe GloS opens up new opportunities for uncertainty quantification in complex systems, allowing for more in depth exploration of model behavior, thereby increasing confidence in model predictions.

Evaluation of Nontargeted Mass Spectral Data Acquisition Strategies for Water Analysis and Toxicity-Based Feature Prioritization by MS2Tox.

The machine-learning tool MS2Tox can prioritize hazardous nontargeted molecular features in environmental waters, by predicting acute fish lethality of unknown molecules based on their MS2 spectra, prior to structural annotation. It has yet to be investigated how the extent of molecular coverage, MS2 spectra quality, and toxicity prediction confidence depend on sample complexity and MS2 data acquisition strategies. We compared two common nontargeted MS2 acquisition strategies with liquid chromatography high-resolution mass spectrometry for structural annotation accuracy by SIRIUS+CSI:FingerID and MS2Tox toxicity prediction of 191 reference chemicals spiked to LC-MS water, groundwater, surface water, and wastewater. Data-dependent acquisition (DDA) resulted in higher rates (19-62%) of correct structural annotations among reference chemicals in all matrices except wastewaters, compared to data-independent acquisition (DIA, 19-50%). However, DIA resulted in higher MS2 detection rates (59-84% DIA, 37-82% DDA), leading to higher true positive rates for spectral library matching, 40-73% compared to 34-72%. DDA resulted in higher MS2Tox toxicity prediction accuracy than DIA, with root-mean-square errors of 0.62 and 0.71 log-mM, respectively. Given the importance of MS2 spectral quality, we introduce a "CombinedConfidence" score to convey relative confidence in MS2Tox predictions and apply this approach to prioritize potentially ecotoxic nontargeted features in environmental waters.

ITER materials irradiation within the D–T neutron environment at JET: post-irradiation radioactivity analysis following the DTE2 experimental campaign

Abstract This work presents the results following the first irradiation of ITER materials samples in a tokamak D--T plasma environment operating at significant fusion power. The materials exposed to this nuclear environment at the Joint European Torus during the DTE2 experimental campaign that took place in 2021 include representative ITER samples from various components such as poloidal field (PF) coil jacket samples, toroidal field coil radial closure plate steels, EUROFER 97 steel, W and CuCrZr materials from the divertor, Inconel-718, CuCrZr, and 316L stainless steel for blanket modules, as well as vacuum vessel forging samples.

The experimental results discussed include high-resolution gamma spectrometry measurements and analysis conducted with the post-irradiated samples, of which there were 68 in total. These samples were exposed through different experimental campaigns, including deuterium, deuterium--tritium and tritium phases. Supporting the analysis were 25 dosimetry foil-based neutron diagnostics and two 'VERDI' neutron spectrometry diagnostics. A further 12 samples for positron annihilation spectroscopy (PALS) were also irradiated. The irradiation of all these samples took place in a long-term irradiation assembly located near the JET vacuum vessel.

The post-irradiation analysis of the ITER material samples has yielded valuable insights into their material activation levels and radiation fields. Comparative assessments between experimental measurements and comprehensive neutronics simulations have demonstrated a significant level of agreement in this work, while also revealing some discrepancies in specific material instances. The data and interpretation from this work not only serve as a robust experimental foundation for enhancing the precision and predictability of neutronics simulation approaches for ITER and next-step devices but also present some opportunities for the refinement of simulation methodologies. In light of these findings, a series of recommendations have been proposed, aimed at improving confidence in nuclear predictions associated with materials that have been exposed to fusion nuclear environments and advancing understanding in this important domain.

Clinical usability of deep learning-based saliency maps for occlusion myocardial infarction identification from the prehospital 12-Lead electrocardiogram

IntroductionDeep learning (DL) models offer improved performance in electrocardiogram (ECG)-based classification over rule-based methods. However, for widespread adoption by clinicians, explainability methods, like saliency maps, are essential. MethodsOn a subset of 100 ECGs from patients with chest pain, we generated saliency maps using a previously validated convolutional neural network for occlusion myocardial infarction (OMI) classification. Three clinicians reviewed ECG-saliency map dyads, first assessing the likelihood of OMI from standard ECGs and then evaluating clinical relevance and helpfulness of the saliency maps, as well as their confidence in the model's predictions. Questions were answered on a Likert scale ranging from +3 (most useful/relevant) to −3 (least useful/relevant). ResultsThe adjudicated accuracy of the three clinicians matched the DL model when considering area under the receiver operating characteristics curve (AUC) and F1 score (AUC 0.855 vs. 0.872, F1 score = 0.789 vs. 0.747). On average, clinicians found saliency maps slightly clinically relevant (0.96 ± 0.92) and slightly helpful (0.66 ± 0.98) in identifying or ruling out OMI but had higher confidence in the model's predictions (1.71 ± 0.56). Clinicians noted that leads I and aVL were often emphasized, even when obvious ST changes were present in other leads. ConclusionIn this clinical usability study, clinicians deemed saliency maps somewhat helpful in enhancing explainability of DL-based ECG models. The spatial convolutional layers across the 12 leads in these models appear to contribute to the discrepancy between ECG segments considered most relevant by clinicians and segments that drove DL model predictions.

College Chatbot Using RASA

The rapid advancement of Artificial Intelligence, Machine Learning, etc., has brought about significant changes in the technological sphere for decades. One such transformative technology is a Chatbot, a Conversational AI capable of addressing users’ queries. These chatbots are tailored to serve various objectives, one of which is to streamline tasks in education through automation. In line with this, a paper titled “College Chatbot Using RASA” has been written, which can effectively respond to questions based on college and other college-related information. The chatbot has been made accessible via an API utilizing the Streamlit framework. The RASA framework, a key open-source technology, has been harnessed to enable the chatbot. RASA plays a vital role in the chatbot's functionality by combining two main components, Rasa Core and Rasa Natural Language Understanding (NLU). Rasa NLU component is used to deduce intent and extract necessary entities from user input, while Rasa Core generates output by constructing a probabilistic model. The model is evaluated by getting a confusion matrix and prediction confidence distribution. The chatbot, upon testing, showcases the best result of 1 for each accuracy, precision, recall, and F1 score.

Development and Validation of a Deep Learning Model for Prediction of Adult Physiological Deterioration.

Prediction-based strategies for physiologic deterioration offer the potential for earlier clinical interventions that improve patient outcomes. Current strategies are limited because they operate on inconsistent definitions of deterioration, attempt to dichotomize a dynamic and progressive phenomenon, and offer poor performance. Can a deep learning deterioration prediction model (Deep Learning Enhanced Triage and Emergency Response for Inpatient Optimization [DETERIO]) based on a consensus definition of deterioration (the Adult Inpatient Decompensation Event [AIDE] criteria) and that approaches deterioration as a state "value-estimation" problem outperform a commercially available deterioration score? The derivation cohort contained retrospective patient data collected from both inpatient services (inpatient) and emergency departments (EDs) of two hospitals within the University of California San Diego Health System. There were 330,729 total patients; 71,735 were inpatient and 258,994 were ED. Of these data, 20% were randomly sampled as a retrospective "testing set." The validation cohort contained temporal patient data. There were 65,898 total patients; 13,750 were inpatient and 52,148 were ED. DETERIO was developed and validated on these data, using the AIDE criteria to generate a composite score. DETERIO's architecture builds upon previous work. DETERIO's prediction performance up to 12 hours before T0 was compared against Epic Deterioration Index (EDI). In the retrospective testing set, DETERIO's area under the receiver operating characteristic curve (AUC) was 0.797 and 0.874 for inpatient and ED subsets, respectively. In the temporal validation cohort, the corresponding AUC were 0.775 and 0.856, respectively. DETERIO outperformed EDI in the inpatient validation cohort (AUC, 0.775 vs. 0.721; p < 0.01) while maintaining superior sensitivity and a comparable rate of false alarms (sensitivity, 45.50% vs. 30.00%; positive predictive value, 20.50% vs. 16.11%). DETERIO demonstrates promise in the viability of a state value-estimation approach for predicting adult physiologic deterioration. It may outperform EDI while offering additional clinical utility in triage and clinician interaction with prediction confidence and explanations. Additional studies are needed to assess generalizability and real-world clinical impact.

Semi-supervised accuracy predictor-based multi-objective neural architecture search

The rise of neural architecture search (NAS) demonstrates the deep exploration between the neural network architecture and its performance (e.g., accuracy). Many NAS methods are inefficient because they train all candidates from scratch to obtain their accuracies. Although predictor-based NAS algorithms have been vigorously developed to efficiently and accurately evaluate the performance of candidate architectures, the training of accuracy predictors still require hundreds of architectures with ground truth. To overcome this shortcoming, this paper investigates an evolutionary-based NAS method, which constructs a semi-supervised accuracy predictor to efficiently and accurately evaluate candidate architectures. A one-time extractor and strong regressors are implemented to further enhance the prediction performance of the semi-supervised accuracy predictor. Furthermore, a multi-objective approach is developed to find architectures with high ground truth in a tradeoff between high prediction accuracy and prediction confidence. Experimental results demonstrate the strong competitiveness of the proposed approach on NAS benchmarks. The code is available at https://github.com/outofstyle/SAPMNAS.

Predicting compound agricultural drought and hot events using a Cascade Modeling framework combining Bayesian Model Averaging ensemble with Vine Copula (CaMBMAViC)

Compound Agricultural Drought and Hot Events (CADHEs) cause more disastrous impacts on water-food-energy-ecology security in comparison with individual drought or hot extremes. Reliable CADHEs predictions are therefore paramount to mitigate the potential risk associated with drought and hot events and their combinations. However, there is no prediction model with multi-predictand variables for reliably predicting CADHEs by considering the compounding influence of diverse confounders. In this study, we first employed a Cascade Modeling (CaM) framework to measure the cascading effects between agricultural droughts and hot events driven by the compounding influence of different confounders. Then, a new ensemble prediction model with multi-predictand variables using a CaM framework coupling Bayesian Model Averaging ensemble with Vine Copula, namely the CaMBMAViC model, was developed to simultaneously predict agricultural droughts, hot events, and CADHEs in the warm season with 1–3-month lead times in China, in which the three most important predictors were selected as the explanatory variables. For these selected typical years (e.g., 2006, 2013, and 2019, as the droughts and hot events in these years swept over most regions of China), the proposed CaMBMAViC model has provided reliable predictions for agricultural droughts, hot events, and CADHEs with 1–3-month lead times in the warm season over most areas of China. Two performance metrics, Kling-Gupta Efficiency (KGE) and Percent Bias (PBIAS), both demonstrated that this model yielded skillful predictions for agricultural droughts (KGE > 0.9 and PBIAS = [1.79%, 1.96%]), hot events (KGE > 0.86 and PBIAS = [2.23%, 2.72%]), and CADHEs (KGE > 0.83 and PBIAS = [–0.95%, 0.36%]) for most areas in China. These findings enhance our confidence in seasonal predictions of compound climate events and help us understand their future dynamic.

A Conformal Prediction Approach to Enhance Predictive Accuracy and Confidence in Machine Learning Application in Chronic Diseases.

Heterogeneity in chronic malignancies raises an increasing interest for the integration and study of predictive models. This study presents a machine learning model approach to predict outcomes and improve their trustworthiness in multi-factorial diseases with highly heterogeneous outcomes, like Chronic Lymphocytic Leukemia (CLL). We incorporated Conformal Prediction to quantify our models uncertainty, and generate confident personalized prediction outcomes that can be integrated into clinical practice.

Damage detection and localization in sealed spent nuclear fuel dry storage canisters using multi-task machine learning classifiers

Spent nuclear fuel (SNF) assemblies (FAs), composed of bundled radioactive fuel rods, are stored in stainless-steel canisters as an interim dry storage option until permanent storage solutions are available. Accidental damage to these canisters may occur during handling or transportation events. If such events occur, it is necessary to assess the integrity of FAs before long-term storage. Since the canisters are sealed shut and can only be opened in special handling facilities, non-destructive evaluation (NDE) is critical for detecting the FAs from the canister's exterior surface. In this study, two multi-task machine learning (ML) classifiers, a k-nearest neighbors (k-NN) and a convolutional neural network (CNN), are developed to simultaneously detect and localize internal FA damage. The classifiers were trained and tested on a dataset collected via experimental modal analysis on a 2/3-scale mock-up canister. The canister was excited from the bottom plate while accelerometers were also attached on the bottom plate at arbitrary locations to record the structural response. The differences in frequency response functions (FRFs) between an intact fully loaded canister basket (FLCB) system and the canister with simulated internal damage were calculated and used as input to the ML models. Results showed that both classifiers achieved high accuracy on the testing set. The k-NN classifier produced macro-F1 scores of 1.00 for the damage detection task and 0.996 for the localization task. The macro-F1 scores of the CNN were 0.991 for damage detection and 0.964 for damage localization. Additionally, dropout layers were added to the fully connected layers of the CNN to introduce model uncertainty. By testing the model 1,000 times, probability density functions (PDFs) were generated, and it was confirmed that the CNN produced confident predictions in both the damage detection and localization tasks. This multi-task ML method contributes to advancing NDE of SNF canisters and holds potential for field applications in inspecting actual SNF canisters.

Simultaneous predictive bands for functional time series using minimum entropy sets

Functional Time Series (FTS) are sequences of dependent random elements taking values on some functional space. Most of the research on this domain focuses on producing a predictor able to forecast the next function, having observed a part of the sequence. For this, the Autoregressive Hilbertian process is a suitable framework. Here, we address the problem of constructing simultaneous predictive confidence bands for a stationary FTS. The method is based on an entropy measure for stochastic processes. To construct predictive bands, we use a Reproducing Kernel Hilbert Space (RKHS) to represent the functions and a functional bootstrap procedure that allows us to estimate the prediction law and a Reproducing Kernel Hilbert Spaces (RKHS) to represent the functions, considering then the basis associated to the reproducing kernel. We then classify the points on the projected space according to those that belong to the minimum entropy set (MES) and those that do not. We map the minimum entropy set back to the functional space and construct a band using the regularity property of the RKHS. The proposed methodology is illustrated through artificial and real data sets.

The OECD (Q)SAR Assessment Framework: A tool for increasing regulatory uptake of computational approaches

There is international interest in using alternatives to animal testing, including (Q)SARs, in chemical hazard assessments. The regulatory acceptance of alternative methods requires principles for considering the scientific rigour of methods and their results. The OECD (Q)SAR assessment Framework (QAF) was developed as guidance for regulators when considering (Q)SAR models and predictions in chemical evaluation. The QAF builds on existing principles for evaluating models and, learning from the longstanding regulatory experience in assessing (Q)SAR predictions, establishes new principles for evaluating predictions and results from multiple predictions. Assessment elements, identified for all principles lay out criteria for assessing the confidence and uncertainties in (Q)SAR models and predictions, while maintaining the flexibility necessary to adapt to different regulatory contexts and purposes. Using the QAF, assessors can consistently and transparently evaluate and decide on the validity of (Q)SARs, and model developers and users have clear requirements to meet. The publication of the QAF is expected to increase the regulatory use and acceptance of (Q)SARs and may become an example to buildsimilar prescriptive frameworks for other new approach methodologies (NAMs). This article provides an overview of the main scientific aspects of the QAF guidance and provides context for how this guidance can promote the use of alternative methods in chemical assessments.

OncoGPT: An AI assistant for genomic-driven precision oncology.

160 Background: Cancer patient’s unique genomic profile can help oncologists select a course of precise personalized treatment for the subject. High throughput next-generation sequencing (NGS) technology allows us to identify genomic alterations simultaneously within a patient’s tumor tissue and blood-circulating tumor DNA (ctDNA) in a single test, however, robust tool for accessing mutation-treatment matching is limited. Methods: OncoGPT was built on a cloud-based elastic computing platform. The NGS data analytical module consists of a cascade of computational algorithms for NGS data processing and gene variant calling. The interactive and univariate Cox proportional hazards models were used for mutation-treatment matching and prognostic effect analysis, respectively. Machine learning algorithms including decision tree, random forest, and neural network were trained and tested with novel features for tissue-of-origin (TO) classification across 8 caner types. Results: We built an AI-driven NGS data analytical platform by integrating computational models, matching algorithms and variant annotation databases to highly accurate achieve cancer-related gene mutations, copy number variation, and structure variants using NGS data from 35,122 tumors across 8 cancer types from three institutions. Next, an AI-driven TO classifier was developed and achieved a weighted F1 score of 0.926 for high confidence predictions (≥ 0.9) on tumor samples. Furthermore, augmented AI matching algorithms were applied to match the optimal personalized treatment and provide prognostic prediction for cancer patients with significantly bett survival outcomes (hazard ratio (HR) = 0.326; 95% confidence interval (CI) = 0.213–0.565; P = 2.52×10−5). Conclusions: We have successfully developed an innovative intelligent system (OncoGPT) with AI capabilities to help accurately find actionable targets from patient tumor or blood ctDNA NGS sequencing data and precisely match individualized therapeutic and clinical trial options for patients. In addition, OncoGPT classified primary tumor sites across 8 different cancer types with high confidence predictions. We believe that OncoGPT would help clinicians make optimal treatment decisions for cancer patients through genomic-driven precision oncology.

Progressive Decision Boundary Shifting for Unsupervised Domain Adaptation.

Unsupervised domain adaptation (UDA) is attracting more attention from researchers for boosting the task-specific generalization on target domain. It focuses on addressing the domain shift between the labeled source domain and the unlabeled target domain. Recent biclassifier-based UDA models perform category-level alignment to reduce domain shift, and meanwhile, self-training is used for improving the discriminability of target instances. However, the error accumulation problem of instances with high semantic uncertainty may cause discriminability degradation and category-level misalignment. To solve this issue, we design the progressive decision boundary shifting algorithm, where stable category information of target instances is explored for learning a discriminability structure on target domain. Specifically, we first model the semantic uncertainty of instances by progressively shifting decision boundaries of category. Then, we introduce the uncertainty decoupling in a contrastive manner, where the discriminative information is learned from the source domain for instance with low semantic uncertainty. Furthermore, we minimize the predictive entropy of instances with high semantic uncertainty to reduce their prediction confidence. Extensive experiments on three popular datasets show that our model outperforms the current state-of-the-art (SOTA) UDA methods.

Assessing allergy risk from ornamental trees in a city: Integrating open access remote sensing data with pollen measurements

Platanus sp. pl. (plane trees) are common ornamental tree in Poland that produces a large amount of wind-transported pollen, which contains proteins that induce allergy symptoms. Allergy sufferers can limit their contact with pollen by avoiding places with high pollen concentrations, which are restricted mainly to areas close to plane trees. Their location is thus important, but creating a detailed street tree inventory is expensive and time-consuming. However, high-resolution remote sensing data provide an opportunity to detect the location of specific plants. But acquiring high-resolution spatial data of good quality also incurs costs and requires regular updates. Therefore, this study explored the potential of using open access remote sensing data to detect plane trees in the highly urbanized environment of Poznań (western Poland). Airborne light detection and ranging (LiDAR) was used to detect training treetops, which were subsequently marked as young plane trees, mature plane trees, other trees or artefacts. Spectral and spatial variables were extracted from circular buffers (r = 1 m) around the treetops to minimize the influence of shadows and crown overlap. A random forest machine learning algorithm was applied to assess the importance of variables and classify the treetops within a radius of 6.2 km around the functioning pollen monitoring station. The model performed well during 10-fold cross-validation (overall accuracy ≈ 92%). The predicted Platanus sp. pl. locations, aggregated according to 16 wind directions, were significantly correlated with the hourly pollen concentrations. Based on the correlation values, we established a threshold of prediction confidence, which allowed us to reduce the fraction of false-positive predictions. We proposed the spatially continuous index of airborne pollen exposure probability, which can be useful for allergy sufferers. The results showed that open-access geodata in Poland can be applied to recognize major local sources of plane pollen.

Delving into human α1,4-galactosyltransferase acceptor specificity: The role of enzyme dimerization

Human α1,4-galactosyltransferase (A4galt), a Golgi apparatus-resident GT, synthesizes Gb3 glycosphingolipid (GSL) and P1 glycotope on glycoproteins (GPs), which are receptors for Shiga toxin types 1 and 2. Despite the significant role of A4galt in glycosylation processes, the molecular mechanisms underlying its varied acceptor specificities remain poorly understood. Here, we attempted to elucidate A4galt specificity towards GSLs and GPs by exploring its interaction with GTs with various acceptor specificities, GP-specific β1,4-galactosyltransferase 1 (B4galt1) and GSL-specific β1,4-galactosyltransferase isoenzymes 5 and 6 (B4galt5 and B4galt6). Using a novel NanoBiT assay, we found that A4galt can form homodimers and heterodimers with B4galt1 and B4galt5 in two cell lines, human embryonic kidney cells (HEK293T) and Chinese hamster ovary cells (CHO-Lec2). We found that A4galt-B4galts heterodimers preferred N-terminally tagged interactions, while in A4galt homodimers, the favored localization of the fused tag depended on the cell line used. Furthermore, by employing AlphaFold for state-of-the-art structural prediction, we analyzed the interactions and structures of these enzyme complexes. Our analysis highlighted that the A4galt-B4galt5 heterodimer exhibited the highest prediction confidence, indicating a significant role of A4galt heterodimerization in determining enzyme specificity toward GSLs and GPs. These findings enhance our knowledge of A4galt acceptor specificity and may contribute to a better comprehension of pathomechanisms of the Shiga toxin-related diseases.

Multi-head co-training: An uncertainty-aware and robust semi-supervised learning framework

Co-training, an advanced form of self-training, allows multiple base models to learn collaboratively, leading to superior performance in semi-supervised learning tasks. However, its widespread adoption is hindered by high computational costs and intricate design choices. To address these challenges, we present Multi-Head Co-Training, a streamlined and efficient framework that consolidates individual models into a multi-head structure, adding minimal extra parameters. Each classification head in this unified model collaborates with others via a “Weak and Strong Augmentation” strategy, with diversity organically introduced through robust data augmentation. Consequently, our approach implicitly promotes diversity while incurring only a minor increase in computational overhead, making co-training more accessible. We validate the effectiveness of Multi-Head Co-Training through an empirical study on standard semi-supervised learning benchmarks. For example, our method achieves up to a 3.1% accuracy improvement on the semi-supervised CIFAR dataset compared to recent methods.Recognizing the necessity for more practical performance metrics beyond accuracy, we assess our framework from three additional perspectives: robust generalization, uncertainty, and computational efficiency. To evaluate robust generalization, we expand the conventional SSL experimental setting to a more comprehensive open-set semi-supervised learning scenario. For uncertainty assessment, we conduct experiments on model calibration and selective classification benchmarks. For example, our method achieves up to a 4.3% accuracy improvement on the open-set semi-supervised CIFAR dataset. Our extensive experiments confirm that our proposed framework better captures prediction confidence and uncertainty, rendering it more suitable for SSL deployment in open environments. The code is available at https://github.com/chenmc1996/Multi-Head-Co-Training.

Explaining transfer learning models for the detection of COVID-19 on X-ray lung images

Amidst the coronavirus disease 2019 (COVID-19) pandemic, researchers are exploring innovative approaches to enhance diagnostic accuracy. One avenue is utilizing deep learning models to analyze lung X-ray images for COVID-19 diagnosis, complementing existing tests like reverse transcription polymerase chain reaction (RT-PCR). However, trusting these models, often viewed as black boxes, presents a challenge. To address this, six explainable artificial intelligence (XAI) techniques: local interpretable model agnostic explanations (LIME), Shapley additive explanations (SHAP), integrated gradients, smooth-grad, gradient-weighted class activation mapping (Grad-CAM), and Layer-CAM are applied to interpret four transfer learning models. These models: VGG16, ResNet50, InceptionV3, and DenseNet121 are analyzed to understand their workings and the rationale behind their predictions. Validating the results with medical experts poses difficulties due to time and resource constraints, alongside the scarcity of annotated X-ray datasets. To address this, a voting mechanism employing different XAI methods across various models is proposed. This approach highlights regions of lung infection, potentially reducing individual model biases stemming from their structures. If successful, this research could pave the way for an automated system for annotating infection regions, bolstering confidence in predictions and aiding in the development of more effective diagnostic tools for COVID-19.

Multi-Task Credible Pseudo-Label Learning for Semi-Supervised Crowd Counting.

As a widely used semi-supervised learning strategy, self-training generates pseudo-labels to alleviate the labor-intensive and time-consuming annotation problems in crowd counting while boosting the model performance with limited labeled data and massive unlabeled data. However, the noise in the pseudo-labels of the density maps greatly hinders the performance of semi-supervised crowd counting. Although auxiliary tasks, e.g., binary segmentation, are utilized to help improve the feature representation learning ability, they are isolated from the main task, i.e., density map regression and the multi-task relationships are totally ignored. To address the above issues, we develop a multi-task credible pseudo-label learning (MTCP) framework for crowd counting, consisting of three multi-task branches, i.e., density regression as the main task, and binary segmentation and confidence prediction as the auxiliary tasks. Multi-task learning is conducted on the labeled data by sharing the same feature extractor for all three tasks and taking multi-task relations into account. To reduce epistemic uncertainty, the labeled data are further expanded, by trimming the labeled data according to the predicted confidence map for low-confidence regions, which can be regarded as an effective data augmentation strategy. For unlabeled data, compared with the existing works that only use the pseudo-labels of binary segmentation, we generate credible pseudo-labels of density maps directly, which can reduce the noise in pseudo-labels and therefore decrease aleatoric uncertainty. Extensive comparisons on four crowd-counting datasets demonstrate the superiority of our proposed model over the competing methods. The code is available at: https://github.com/ljq2000/MTCP.