LungDetectNet: a multi-task deep learning framework with enhanced detection and descriptive capabilities

The social-economic statistics collected from local governments are the main access for the central government to achieve national economic circumstance, especially for China. However, the statistics of almost 10% of national counties are missing or inconsistent due to the statistical caliber change in the wave of urbanization during economic development. Some researchers proposed to apply a night luminosity product to solve such issue. However, it lacks the ability to distinguish between the wealthy populations with a dense distribution and the less developed places. In this paper, the publicly available daytime Landsat images are used to estimate economic statistics. An end-to-end multi-task deep learning framework is constructed to estimate the county-level economy of Mainland China and the overall accuracy of this model achieves higher than 85%. The experiments show that our model provides a potential strategy to make up for the missing statistics and examines the reliability of the statistics collected for the central government.

The human iris contains rich physiological information through its texture, color, and morphology. Traditional iris analysis methods are limited by computational inefficiency and poor scalability for multi-indicator prediction. This study aimed to develop and validate a lightweight, multi-task deep learning framework that can noninvasively predict multiple physiological indicators from high-resolution iris images. We developed a hybrid model combining ConvNeXt-Tiny and TinyViT backbones, trained on 30,353 iris images with a multi-task learning (MTL) strategy. To enhance performance with limited labeled data, semi-supervised learning with consistency regularization and pseudo-labeling was integrated. Fast Adaptive Multi-Task Optimization (FAMO) and Nash Multi-Task Learning (Nash-MTL) were used to balance task-specific losses. Feature vectors with 2400 dimensions — derived from texture, chromatic, and morphological cues — were compressed to 120 dimensions using Principal Component Analysis (PCA), preserving 96.2% of the data variance. This dimensionality reduction led to a 2.9-fold improvement in inference speed and reduced feature storage size by nearly 80% significantly enhancing computational efficiency without compromising predictive accuracy. The model achieved 89.1% average accuracy across 13 physiological tasks, with F1 scores of 0.91 for heart and 0.89 for brain indicators, outperforming fully supervised baselines. Among these, 7 organ-related indicators were validated using data from 60 Parkinson’s disease patients, showing strong left–right eye correlation ([Formula: see text], [Formula: see text]).

In this paper, we propose an efficient method to segment the leaves and count the number of leaves in digital plant images, important tasks in plant phenotyping. However, for large cropping lands, automated leaf counting and segmentation are challenging and time-consuming due to the plant’s rapid growth, varying size of leaves and shape, and presence of occlusions, which pose the task challenging to even human experts. Deep learning methods show considerable improvement when employed in image-based plant phenotyping tasks like plant leaf counting and leaf segmentation. However, most of the research related to plant phenotyping was designed to perform a single-task at a time. Therefore, a multi-task deep learning (MTL) framework is proposed to infer the two tasks simultaneously (i)leaf segmentation and (ii)leaf counting. Compared to single-task learning, MTL achieved the segmentation mean intersection over union score of 0.7478 (improved by 2%), and the difference in the count is decreased by two times when employed in agricultural field images.

Adverse Drug Reactions (ADRs) have a direct impact on human health. As continuous pharmacovigilance and drug monitoring prove to be costly and time-consuming, computational methods have emerged as promising alternatives. However, most existing computational methods primarily focus on predicting whether or not the drug is associated with an adverse reaction and do not consider the core issue of drug benefit-risk assessment—whether the treatment outcome is serious when adverse drug reactions occur. To this end, we categorize serious clinical outcomes caused by adverse reactions to drugs into seven distinct classes and present a deep learning framework, so-called GCAP, for predicting the seriousness of clinical outcomes of adverse reactions to drugs. GCAP has two tasks: one is to predict whether adverse reactions to drugs cause serious clinical outcomes, and the other is to infer the corresponding classes of serious clinical outcomes. Experimental results demonstrate that our method is a powerful and robust framework with high extendibility. GCAP can serve as a useful tool to successfully address the challenge of predicting the seriousness of clinical outcomes stemming from adverse reactions to drugs.

To develop a comprehensive deep learning and radiomics-based multi-task pipeline for the detection and classification of key fetal anatomical structures in first-trimester ultrasound images, using a diverse multi-center dataset to ensure high variability, reproducibility, and generalizability. A total of 4,532 fetal ultrasound scans (gestational age 11–14 weeks), retrospectively collected from nine medical centers, were included in this study. Two detection models, You Only Look Once version 11 (YOLOv11) and shifted window transformer (Swin Transformer), were trained to localize nine fetal brain and craniofacial structures. From each detected region, 215 radiomic features and 1,792 deep features were extracted. Feature stability was ensured through intra-class correlation coefficient (ICC) filtering (threshold ≥ 0.75), Pearson correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. The selected features were then used to train a Transformer-based model for tabular data (TabTransformer) to classify fetal anatomical structures into clinically defined categories based on their sonographic appearance. Model performance was evaluated using Accuracy, area under the receiver operating characteristic curve (AUC), and Sensitivity across training, internal validation, and external test datasets. Fusion models integrating radiomic and deep features consistently outperformed single-modality models in both detection and classification. On the external test set, classification accuracy reached 96.1%, with AUCs up to 96.89%, and sensitivity exceeding 95% for key anatomical structures. Swin Transformer achieved superior localization performance compared to YOLOv11, with Intersection over Union (IoU) values up to 0.97 and F1-scores ≥ 0.94. Feature reproducibility remained above 75% across centers. The TabTransformer classifier demonstrated strong generalization and robustness, effectively leveraging the fused feature space for high-precision classification. This study presents the fully integrated, multi-task framework for fetal anatomical structure detection and classification using multi-center ultrasound data. The proposed approach demonstrates high reproducibility and diagnostic performance, offering strong clinical potential for early and objective fetal anomaly screening in the first trimester.

Colorectal cancer (CRC) is the third most commonly diagnosed cancer and the second most lethal cancer in the world. The segmentation and staging of tumours in CT images are important for CRC diagnosis and treatment. Using deep learning to fulfil medical image diagnosis is a promising approach. Most current methods solve segmentation and staging tasks separately, while ignoring the associated information between tasks. Multi-task deep learning (MTL) can solve the two tasks simultaneously. In this study, we proposed a CRC_MTL framework applying multi-task deep learning to complete two tasks jointly for analysing CRC tumours in abdominal CT images. The CRC_MTL framework not only improves the accuracy of the classification task from 91.40% to 93.52% but also improves the dice coefficient of the segmentation task from 0.6900 to 0.7379.

In recent years, convolutional neural networks (CNNs) have made great achievements in object extraction from very high-resolution (VHR) images. However, most existing approaches require large quantities of clean and accurate training data to achieve impressive classification results. The presence of inaccurate labels in training datasets is known to deteriorate the performance of CNNs. In this paper, we introduce a novel efficient method for improving the robustness when training CNN on the dataset with relatively noisy labels. First, we propose a feature and label noise model (FLNM) to model the noisy label distribution in the training dataset. Then, we use a multitask deep learning framework (MDLF) to integrate the FLNM into the training process of CNN. Finally, a novel loss function concerning the high-level features is introduced to efficiently train the MDLF. We evaluate our method on datasets from Massachusetts and compare this method with other state-of-the-art methods. The experimental results demonstrate the effectiveness of the proposed method in improving the classification performance of CNNs trained with noisy training dataset.

Due to the potentially significant benefits for society, forecasting spatio-temporal societal events is currently attracting considerable attention from researchers. Beyond merely predicting the occurrence of future events, practitioners are now looking for information about specific subtypes of future events in order to allocate appropriate amounts and types of resources to manage such events and any associated social risks. However, forecasting event subtypes is far more complex than merely extending binary prediction to cover multiple classes, as 1) different locations require different models to handle their characteristic event subtype patterns due to spatial heterogeneity; 2) historically, many locations have only experienced a incomplete set of event subtypes, thus limiting the local model’s ability to predict previously “unseen” subtypes; and 3) the subtle discrepancy among different event subtypes requires more discriminative and profound representations of societal events. In order to address all these challenges concurrently, we propose a Spatial Incomplete Multi-task Deep leArning (SIMDA) framework that is capable of effectively forecasting the subtypes of future events. The new framework formulates spatial locations into tasks to handle spatial heterogeneity in event subtypes, and learns a joint deep representation of subtypes across tasks. Furthermore, based on the “first law of geography”, spatiallyclosed tasks share similar event subtype patterns such that adjacent tasks can share knowledge with each other effectively. Optimizing the proposed model amounts to a new nonconvex and strongly-coupled problem, we propose a new algorithm based on Alternating Direction Method of Multipliers (ADMM) that can decompose the complex problem into subproblems that can be solved efficiently. Extensive experiments on six real-world datasets demonstrate the effectiveness and efficiency of the proposed model.

Deep reinforcement learning (RL) approaches are increasingly prominent in the field of robotics due to their adaptive decision-making capability. However, developing a single RL agent capable of performing multiple continuous control tasks for quadruped robots remains challenging. In this paper, a multi-task deep RL framework based on curriculum learning and policy distillation is proposed, which aims to enhance the quadruped robot's motor performance across multiple continuous tasks by leveraging knowledge from expert skill teachers. The main novelties of the framework lie in the self-optimizing terrain curriculum learning strategy and the improved distillation loss function. The proposed self-optimizing terrain curriculum learning strategy for quadrupedal robots is designed to utilize Bayesian optimization to predict potential training terrains, thus effectively identifying the most suitable training curriculum. Additionally, the improved distillation loss function for RL weight optimization is proposed to enhance the transferability of the trained policy across diverse tasks. To validate the effectiveness of the proposed multi-task deep RL framework, the performance of the policy generated by the framework across diverse terrains is assessed. The experimental results demonstrate that the proposed multi-task deep RL framework could generate a unified policy that achieves excellent performance across multiple continuous control tasks for quadruped robots.

Purpose To develop and evaluate a novel multitask deep learning framework for automated detection and localization of endoleaks at aortic digital subtraction angiography (DSA) performed during real-world endovascular aneurysm repair (EVAR) procedures for abdominal aortic aneurysm. Materials and Methods This retrospective study analyzed intraoperative aortic DSA images from patients undergoing EVAR (January 2017-December 2021). An expert panel assessed each sequence for endoleaks. Each sequence was processed into three input channels: peak density, time to peak, and area under the time-density curve, generating three two-dimensional perfusion maps per patient. These maps served as input into a convolutional neural network for binary detection (classification) and localization (regression) of endoleaks through multitask learning. Fivefold cross-validation was performed, with patients split 80:20 into training and testing datasets for each fold. Performance metrics included area under the receiver operating characteristic curve, F1 score, precision, and recall and were compared with human experts. Results The study included 220 patients (median age, 74 years [IQR, 68-79]; 181 male). Endoleaks were visible in 111 of 220 (50.5%) patients. The model identified and localized endoleaks with an area under the receiver operating characteristic curve of 0.85 ± 0.0031 (SD), F1 score of 0.78 ± 0.21, 95% precision, and 73% recall. Compared with the procedural team (94% precision, 63% recall), it had higher values in both metrics, with an F1 score within the human observer range (0.75-0.85). Balancing regression and classification by multitask learning delivered optimal results. The interobserver agreement among human experts was moderate (Fleiss κ = 0.404). Conclusion A novel, fully automated deep learning method accurately detected and localized endoleaks at DSA imaging from EVAR procedures. Keywords: Endoleaks, Deep Learning, Digital Subtraction Angiography, Endovascular Aneurysm Repair, Abdominal Aortic Aneurysm, Interventional Suite, Hybrid Operating Room Supplemental material is available for this article. © RSNA, 2025.

A unified modeling framework for lane change intention recognition and vehicle status prediction

Retinal diseases represent a leading cause of visual impairment and blindness worldwide, with early and accurate diagnosis being crucial for preventing irreversible vision loss. Although deep learning techniques have achieved significant advances in fundus image analysis, existing methods predominantly focus on single tasks, treating vessel segmentation and disease diagnosis as independent problems without fully leveraging their intrinsic relationships. Furthermore, the lack of transparency in deep model predictions limits clinical adoption; while full interpretability remains an open challenge, post-hoc techniques such as Grad-CAM can provide partial transparency by highlighting influential image regions. This study presents V-MNet, a vessel-guided multi-task deep learning framework that simultaneously achieves retinal vessel segmentation and multi-disease classification, and provides visual transparency through Grad-CAM-based class activation mapping to support clinical decision-making. The framework comprises three core modules: a shared encoder extracts multi-scale feature representations; a segmentation decoder employs a U-Net-style architecture to generate vessel masks; and a classification decoder incorporates an innovative vessel-guided mechanism that explicitly transfers structural priors from the segmentation branch to the classification task, enabling the model to precisely localize pathological regions. Concurrently, an integrated Grad-CAM module generates post-hoc class activation maps for each disease category, highlighting spatially relevant lesion regions for clinician review. Comprehensive experiments were conducted on four public datasets-RFMiD, ODIR-5K, DRIVE, and EyePACS-light-v2. Experiments demonstrate that V-MNet achieves a Dice coefficient of 0.831 and AUC of 0.985 for vessel segmentation tasks, and an average AUC of 0.978 with F1-score of 0.935 for multi-disease classification tasks, significantly outperforming single-task baseline models and existing state-of-the-art methods. Ablation studies systematically quantify the performance contributions of multi-task learning and the vessel-guided mechanism, confirming the effectiveness of the framework's core innovations. V-MNet demonstrates broad application potential as a computer-aided diagnostic tool by jointly leveraging vascular structure and disease pathology for superior performance and visual transparency. The vessel-guided multi-task design effectively exploits the intrinsic relationship between vessel segmentation and disease classification, while the integrated Grad-CAM module addresses the lack of model transparency, facilitating clinical adoption and supporting clinical decision-making.

UltraMN: Advancing Real-Time Median Nerve Ultrasound Monitoring With a Multitask Deep Learning Framework.

Trajectory regression, which aims to predict the travel time of arbitrary trajectories on road networks, attracts significant attention in various applications of traffic systems these years. In this paper, we tackle this problem with a multitask learning (MTL) framework. To take the temporal nature of the problem into consideration, we divide the regression problem into a set of sub-tasks of distinct time periods, then the problem can be treated in a multi-task learning framework. Further, we propose a novel regularization term in which we exploit the block sparse structure to augment the robustness of the model. In addition, we incorporate the spatial smoothness over road links and thus achieve a spatial-temporal framework. An accelerated proximal algorithm is adopted to solve the convex but non-smooth problem, which will converge to the global optimum. Experiments on both synthetic and real data sets demonstrate the effectiveness of the proposed method.

In this work we utilize implied volatility predictions from options on individual securities to predict the average volatility of an ensemble of securities. We cast this as a problem of online learning with expert advice, which has been studied extensively in the Machine Learning community. In this setting we treat each security as an expert and adaptively maintain a set of weights to combine the implied volatility predictions of all securities in our ensemble. We also consider options of different term lengths, treating each term length as a separate learning task in a Multi-task Learning (MTL) framework. In this MTL framework information is shared between the learning tasks to improve the overall learning process. We conduct experiments using historical option pricing data and corresponding historical volatility records. Using the straight mean of the implied volatility predictions as a benchmark strategy, we demonstrate improved combined predictions with a basic online learning algorithm and further improvement with our MTL framework.