Unified Deep Learning Framework with Two-Way Transfer Learning for Occluded Faces: A Soft Decision Fusion Approach

Objective: To design a deep transfer learning framework for modeling fluence map predictions for stereotactic body radiation therapy (SBRT) of adrenal cancer and similar sites that usually have a small number of cases. Approach: We developed a transfer learning framework for adrenal SBRT planning that leverages knowledge in a pancreas SBRT planning model. Treatment plans from the two sites had different dose prescriptions and beam settings but both prioritized gastrointestinal sparing. A base framework was first trained with 100 pancreas cases. This framework consists of two convolutional neural networks (CNN), which predict individual beam doses (BD-CNN) and fluence maps (FM-CNN) sequentially for 9-beam intensity-modulated radiation therapy (IMRT) plans. Forty-five adrenal plans were split into training/validation/test sets with the ratio of 20/10/15. The base BD-CNN was re-trained with transfer learning using 5/10/15/20 adrenal training cases to produce multiple candidate adrenal BD-CNN models. The base FM-CNN was directly used for adrenal cases. The deep learning (DL) plans were evaluated by several clinically relevant dosimetric endpoints, producing a percentage score relative to the clinical plans. Main results: Transfer learning significantly reduced the number of training cases and training time needed to train such a DL framework. The adrenal transfer learning model trained with 5/10/15/20 cases achieved validation plan scores of 85.4/91.2/90.7/89.4, suggesting that model performance saturated with 10 training cases. Meanwhile, a model using all 20 adrenal training cases without transfer learning only scored 80.5. For the final test set, the 5/10/15/20-case models achieved scores of 73.5/75.3/78.9/83.3. Significance: It is feasible to use deep transfer learning to train an IMRT fluence prediction framework. This technique could adapt to different dose prescriptions and beam configurations. This framework potentially enables DL modeling for clinical sites that have a limited dataset, either due to few cases or due to rapid technology evolution.

In defense and security applications, detection of moving target direction is as important as the target detection and/or target classification. In this study, a methodology for the detection of different mobile targets as approaching or receding was proposed for ground surveillance radar data, and convolutional neural networks (CNN) based on transfer learning were employed for this purpose. In order to improve the classification performance, the use of two key concepts, namely Deep Convolutional Generative Adversarial Network (DCGAN) and decision fusion, has been proposed. With DCGAN, the number of limited available data used for training was increased, thus creating a bigger training dataset with identical distribution to the original data for both moving directions. This generated synthetic data was then used along with the original training data to train three different pre-trained deep convolutional networks. Finally, the classification results obtained from these networks were combined with decision fusion approach. In order to evaluate the performance of the proposed method, publicly available RadEch dataset consisting of eight ground target classes was utilized. Based on the experimental results, it was observed that the combined use of the proposed DCGAN and decision fusion methods increased the detection accuracy of moving target for person, vehicle, group of person and all target groups, by 13.63%, 10.01%, 14.82% and 8.62%, respectively.

Prediction of exons locations in deoxyribonucleic acid (DNA) sequences is a significant issue for biologists. This paper proposes a new method to solve this problem. Unlike the published studies, in which the prediction of exons locations depends on hard decisions from a single classifier, the proposed prediction approach depends on fusion of soft decisions from two classifiers. In the proposed approach we utilize the sliding window discrete Fourier transform (DFT), which is normally used to detect exons 3-base periodicity feature, in a different manner. The novelty here depends on obtaining soft decisions, rather than hard decisions, from two classifiers using different numerical mapping schemes, and fuses them in a decision fusion center to obtain a final global decision about the prediction of exons locations. Simulation results based on real data performed on the HMR195 dataset showed that the proposed soft decisions fusion method achieves better prediction performance compared to the traditional hard decision single classifier method. Moreover the proposed method can easily be extended to more than two classifiers.

A comprehensive empirical study on bug characteristics of deep learning frameworks

Deep learning (DL) frameworks serve as the backbone for a wide range of artificial intelligence applications. However, bugs within DL frameworks can cascade into critical issues in higher-level applications, jeopardizing reliability and security. While numerous techniques have been proposed to detect bugs in DL frameworks, research exploring common API patterns across frameworks and the potential risks they entail remains limited. Notably, many DL frameworks expose similar APIs with overlapping input parameters and functionalities, rendering them vulnerable to shared bugs, where a flaw in one API may extend to analogous APIs in other frameworks. To address this challenge, we propose MirrorFuzz, an automated API fuzzing solution to discover shared bugs in DL frameworks. MirrorFuzz operates in three stages: First, MirrorFuzz collects historical bug data for each API within a DL framework to identify potentially buggy APIs. Second, it matches each buggy API in a specific framework with similar APIs within and across other DL frameworks. Third, it employs large language models (LLMs) to synthesize code for the API under test, leveraging the historical bug data of similar APIs to trigger analogous bugs across APIs. We implement MirrorFuzz and evaluate it on four popular DL frameworks (TensorFlow, PyTorch, OneFlow, and Jittor). Extensive evaluation demonstrates that MirrorFuzz improves code coverage by 39.92% and 98.20% compared to state-of-the-art methods on TensorFlow and PyTorch, respectively. Moreover, MirrorFuzz discovers 315 bugs, 262 of which are newly found, and 80 bugs are fixed, with 52 of these bugs assigned CNVD IDs.

Accurate prediction and confirmation of mechanical properties are critical for the reliable design and deployment of sustainable aluminum–agro-waste composites (AWCs). This study presents a confirmatory analytical–experimental–deep machine learning (DML) framework for the identification of true material properties under multi-response conditions. Experimental characterization yielded mean tensile strength of 190.61 MPa, compressive strength of 287.16 MPa, flexural strength of 243.94 MPa, wear rate of 5.46 × 10⁻⁶ mm³ /N·m, coefficient of friction of 0.34, and fatigue life of 5.86 × 10⁵ cycles, with low variability observed in strength-related properties and higher dispersion in tribological and fatigue responses. Isolated analytical predictions produced comparable estimates for tensile (189.60 MPa), compressive (296.42 MPa), flexural (255.75 MPa), and fatigue properties (5.97 × 10⁵ cycles), but failed to reliably capture tribological behavior, yielding a non-physical negative wear rate (− 4.06 mm³/N·m) and an overestimated coefficient of friction (0.45). Consolidated analytical modeling improved agreement for tensile (187.60 MPa), compressive (285.63 MPa), flexural (244.88 MPa), and fatigue life (5.86 × 10⁵ cycles), yet significantly deviated in tribological predictions, producing unrealistically high wear rate (0.20 mm³/N·m) and coefficient of friction (2.84). To resolve these discrepancies, DML models—implemented via a deep convolutional neural network (ID-CNN) and a hybrid residual CNN guided by a residual-based loss function—were employed to reconcile analytical (isolated/consolidated) predictions with experimental data. The DML framework successfully minimized analytical–experimental residuals and confirmed true predictive reliability for compressive strength, flexural strength, coefficient of friction, and fatigue life, while consistently identifying wear rate as falsely predicted across analytical frameworks. These findings establish DML as both a predictive and diagnostic tool, enabling discrimination between truly confirmed and analytically unstable properties. The proposed framework offers a scalable, data-informed pathway for optimized composite design, process refinement, and industrial adoption, with potential extension to multi-scale modeling and real-time process monitoring. Residual-based DML loss quantifies and minimizes analytical–experimental prediction errors. • Residual-based DML loss quantifies and minimizes analytical–experimental prediction errors. • Integrated analytical–DML modeling confirms strength, fatigue, and friction properties. • Bulk properties converge, while wear rate shows persistent tribological uncertainty. • The framework supports reliable certification and sustainable composite design.

In cold climates, ice formation on wind turbines causes power reduction produced by a wind farm. This paper introduces a framework to predict icing at the farm level based on our recently developed Temporal Convolutional Network prediction model for a single turbine using SCADA data.First, a cross-validation study is carried out to evaluate the extent predictors trained on a single turbine of a wind farm can be used to predict icing on the other turbines of a wind farm. This fusion approach combines multiple turbines, thereby providing predictions at the wind farm level. This study shows that such a fusion approach improves prediction accuracy and decreases fluctuations across different prediction horizons when compared with single-turbine prediction. Two approaches are considered to conduct farm-level icing prediction: decision fusion and feature fusion. In decision fusion, icing prediction decisions from individual turbines are combined in a majority voting manner. In feature fusion, features of individual turbines are averaged first before conducting prediction. The results obtained indicate that both the decision fusion and feature fusion approaches generate farm-level icing prediction accuracies that are 7% higher with lower standard deviations or fluctuations across different prediction horizons when compared with predictions for a single turbine.

Deep learning (DL) frameworks are essential to DL-based software systems, and framework bugs may lead to substantial disasters, thus requiring effective testing. Researchers adopt DL models or single interfaces as test inputs and analyze their execution results to detect bugs. However, floating-point errors, inherent randomness, and the complexity of test inputs make it challenging to analyze execution results effectively, leading to existing methods suffering from a lack of suitable test oracles. Some researchers utilize metamorphic testing to tackle this challenge. They design Metamorphic Relations (MRs) based on input data and parameter settings of a single framework interface to generate equivalent test inputs, ensuring consistent execution results between original and generated test inputs. Despite their promising effectiveness, they still face certain limitations. (1) Existing MRs overlook structural complexity, limiting test input diversity. (2) Existing MRs focus on limited interfaces, which limits generalization and necessitates additional adaptations. (3) Their detected bugs are related to the result consistency of single interfaces and far from those exposed in multi-interface combinations and runtime metrics (e.g., resource usage). To address these limitations, we propose ModelMeta, a model-level metamorphic testing method for DL frameworks with four MRs focused on the structure characteristics of DL models. ModelMeta augments seed models with diverse interface combinations to generate test inputs with consistent outputs, guided by the QR-DQN strategy. It then detects bugs through fine-grained analysis of training loss/gradients, memory/GPU usage, and execution time. We evaluate the effectiveness of ModelMeta on three popular DL frameworks (i.e., MindSpore, PyTorch, and ONNX) with 17 DL models from ten real-world tasks ranging from image classification to object detection. Results demonstrate that ModelMeta outperforms state-of-the-art baselines from the perspective of test coverage and diversity of generated test inputs. Regarding bug detection, ModelMeta has identified 31 new bugs, of which 27 have been confirmed, and 11 have been fixed. Among them, seven bugs existing methods cannot detect, i.e., five wrong resource usage bugs and two low-efficiency bugs. These results demonstrate the practicality of our method.

Image-based 3D reconstruction and permeability modelling of rock using enhanced interpretable deep residual learning

BackgroundNasopharyngeal carcinoma (NPC), an endemic malignancy in Southeast Asia and southern China, necessitates precise delineation of the gross tumor volume (GTV) for radiotherapy to optimize patient prognosis. Positron emission tomography (PET) imaging offers valuable metabolic insights to guide radiotherapy planning but is hampered by manual GTV segmentation, which is labor-intensive and prone to inter- and intra-observer variability. Conventional segmentation methods are inherently limited, and comparative evaluations of deep learning (DL) frameworks for PET-based primary NPC lesion segmentation remain scarce. Thus, this study aimed to optimize the delineation of primary NPC radiotherapy target volumes by leveraging PET-image-based DL segmentation technology, with the goal of enhancing both the precision and efficiency of lesion delineation.MethodsResearchers retrospectively collected PET imaging data from 212 NPC patients at the Affiliated Hospital of Guangdong Medical University. The patients were randomly divided into training (170 patients) and testing (42 patients) datasets. A radiation oncologist and a nuclear medicine physician collaboratively delineated the cancer lesion boundaries through consensus. After data preprocessing, three DL models (Res-Unet, Nn-Unet, and Nn-Former) were used to automatically segment the lesions. Training was based on the training dataset. During evaluation, the models’ segmentation performance was comprehensively assessed using Dice similarity coefficient (DSC) and Hausdorff distance (HD). A visual analysis of the results was also conducted to intuitively understand the models’ segmentation capabilities.ResultsIn the training set, Nn-Unet achieved the highest DSC of 0.869, whereas in the testing set, its DSC was 0.833. The DSC for Nn-Former hovered around 0.8 in both the training and testing sets. In contrast, Res-Unet demonstrated the lowest DSC values among the three models, particularly in the testing set (DSC =0.794). Statistical analysis revealed that the DSC of Nn-Unet was significantly higher than that of both Res-Unet and Nn-Former in the training set (P<0.01 for both comparisons). Regarding the 95% HD and average surface distance (ASD) values, Nn-Unet outperformed the other two models in both the training and testing sets. However, Res-Unet exhibited a much higher HD value in the testing set compared to the training set. The Loss curves for all models gradually decreased as training progressed, indicating that the models were learning relevant features of NPC lesions. The final Loss value for Res-Unet was approximately 0.7, whereas Nn-Unet and Nn-Former had final Loss values below 0.8.ConclusionsThe PET-image-based automatic segmentation model for the primary tumor of NPC established in this study demonstrates the clinical potential of Nn-Unet in primary lesion of NPC segmentation tasks.

Traditional differential expression genes (DEGs) identification models have limitations in small sample size datasets because they require meeting distribution assumptions, otherwise resulting high false positive/negative rates due to sample variation. In contrast, tabular data model based on deep learning (DL) frameworks do not need to consider the data distribution types and sample variation. However, applying DL to RNA-Seq data is still a challenge due to the lack of proper labeling and the small sample size compared to the number of genes. Data augmentation (DA) extracts data features using different methods and procedures, which can significantly increase complementary pseudo-values from limited data without significant additional cost. Based on this, we combine DA and DL framework-based tabular data model, propose a model TabDEG, to predict DEGs and their up-regulation/down-regulation directions from gene expression data obtained from the Cancer Genome Atlas database. Compared to five counterpart methods, TabDEG has high sensitivity and low misclassification rates. Experiment shows that TabDEG is robust and effective in enhancing data features to facilitate classification of high-dimensional small sample size datasets and validates that TabDEG-predicted DEGs are mapped to important gene ontology terms and pathways associated with cancer.

Cooperative spectrum sensing (CSS) has the ability to accurately identify the activities of the primary users (PUs). As the secondary users’ (SUs) sensing performance is disturbed in the fading and shadowing environment, therefore the CSS is a suitable choice to achieve better sensing results compared to individual sensing. One of the problems in the CSS occurs due to the participation of malicious users (MUs) that report false sensing data to the fusion center (FC) to misguide the FC’s decision about the PUs’ activity. Out of the different categories of MUs, Always Yes (AY), Always No (AN), Always Opposite (AO) and Random Opposite (RO) are of high interest these days in the literature. Recently, high sensing performance for the CSS can be achieved using machine learning techniques. In this paper, boosted trees algorithm (BTA) has been proposed for obtaining reliable identification of the PU channel, where the SUs can access the PU channel opportunistically with minimum disturbances to the licensee. The proposed BTA mitigates the spectrum sensing data falsification (SSDF) effects of the AY, AN, AO and RO categories of the MUs. BTA is an ensemble method for solving spectrum sensing problems using different classifiers. It boosts the performance of some weak classifiers in the combination by giving higher weights to the weak classifiers’ sensing decisions. Simulation results verify the performance improvement by the proposed algorithm compared to the existing techniques such as genetic algorithm soft decision fusion (GASDF), particle swarm optimization soft decision fusion (PSOSDF), maximum gain combination soft decision fusion (MGCSDF) and count hard decision fusion (CHDF). The experimental setup is conducted at different levels of the signal-to-noise ratios (SNRs), total number of cooperative users and sensing samples that show minimum error probability results for the proposed scheme.

Minimizing waste entering US landfills has become increasingly difficult. Composting organic waste can assist in reducing the effects of landfill pollution and greenhouse gas emissions. To help automate this process, we apply computer vision techniques to train a classifier to distinguish between green and brown compost. The lack of labeled compost data and the challenges and dangers posed by collecting such data create a problem for the automation of composting. In this study, we propose CGBNet, a deep learning framework to differentiate between green and brown compost for the creation of healthy compost piles with an effective ratio of carbon to nitrogen. First, superclass and subclass classification are proposed as two methods of labeling and classifying data. Six deep learning models are then trained and evaluated on a newly formed dataset of 1,960 images using both label structures. Next, transfer learning is applied for feature extraction to boost performance using state-of-the-art deep learning models. The models showed improvements in performance generalization with less overfitting. The best-performing model, CGBNet, used transfer learning and subclass classification to obtain an accuracy of 95%, a 16% improvement in accuracy from the best performing superclass classification model trained without transfer learning. Superclass classification also showed significant improvements in performance and practicality with transfer learning, allowing for robust compost recognition systems to be more extensible while requiring less data and fewer parameters to train.

Supervised machine learning is a powerful and widely used method for analyzing high-content screening data. Despite its accuracy, efficiency, and versatility, supervised machine learning has drawbacks, most notably its dependence on a priori knowledge of expected phenotypes and time-consuming classifier training. We provide a solution to these limitations with CellCognition Explorer, a generic novelty detection and deep learning framework. Application to several large-scale screening data sets on nuclear and mitotic cell morphologies demonstrates that CellCognition Explorer enables discovery of rare phenotypes without user training, which has broad implications for improved assay development in high-content screening.

In this study, a new method for the analysis of harmonic components in the power system based on a deep learning (DL) framework is introduced. In the proposed method, both amplitudes and phases of the harmonic components can be estimated accurately, unlike most of the research work in the literature, which usually focus on estimating amplitudes only. A convolutional neural network (CNN) structure is used to estimate the phases and amplitudes of harmonics, although CNN is usually used for classification. It has been shown that the proposed DL‐based method can satisfactorily estimate both amplitudes and phases of the power system harmonics inside a 20‐ms window and this makes the proposed method suitable for possible real‐time applications, such as active power filtering of the harmonics. It has also been shown that the proposed method is robust to fundamental frequency changes. Experiments on carefully‐generated data sets to reflect the power system behaviour show that the proposed method demonstrates remarkably good performance in terms of estimation accuracy, especially for time‐varying frequency cases. Average error for the amplitude estimation is obtained as 0.21% and that for the phase is 9°, which outperforms the other compared analyses methods in cases of fundamental frequency variations.