Umeå University's logo

The increasing adoption of deep learning in distributed environments is driving the need for efficient deployment strategies across the cloud–edge continuum. Traditional monolithic models, typically executed in centralized cloud infrastructures, often suffer from high latency, excessive bandwidth consumption, and limited support for privacy-sensitive applications. In this paper, we propose a distributed deep learning framework designed according to the Pipes and Filters architectural pattern, where the inference workflow is decomposed into independent processing stages connected through well-defined intermediate representations. To illustrate the applicability of the proposed approach, we outline a representative scenario: a distributed e-health system, where privacy-aware processing is critical. Finally, we discuss possible work toward the evaluation of the architecture through cloud-edge simulation tools, with the goal of quantitatively assessing performance improvements and guiding future optimization strategies.

Agriculture is increasingly challenged by climate variability, resource scarcity, and the need for more efficient and sustainable farm management. Although artificial intelligence has shown strong potential in tasks such as crop monitoring, disease diagnosis, irrigation support, and field robotics, most current systems remain task-specific and operate as isolated analytical or decision-support modules. This paper presents a prospective view of agentic AI in agriculture as an emerging systems paradigm for automatic farming. We distinguish AI-assisted agriculture from agentic agricultural systems, organize recent literature into four directions, and propose a framework for automatic farming systems centered on multimodal perception, specialized decision components, orchestration, bounded actuation, and continuous feedback under human oversight. We also discuss the main challenges for real-world deployment, including partial observability, heterogeneity, multi-objective decision-making, safety, and evaluation. The paper argues that the next step for agricultural AI is not only improved task-level performance, but the design of coherent farm-scale systems able to perceive, reason, and act under real operational constraints.

Background: Thymic epithelial tumors (TETs) are rare mediastinal malignancies whose prognosis is largely determined by histology. Current predictive models rely on clinical variables and subjective imaging interpretation, with unsatisfied performance. Non-invasive pre-treatment risk stratification could guide surgical planning and perioperative management in patients with TETs. The role of fluorine-18 (18F) fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (CT) in identifying aggressive disease is increasingly recognized. In this bicentric study, we aimed to evaluate a machine learning-based radiomics model using PET and CT images to differentiate between low-risk and high-risk TETs.

Methods: Seventy-five patients who underwent PET/CT to evaluate the suspected anterior mediastinal mass and histopathologically diagnosed with TETs were included. On PET/CT images, the tumor was manually segmented by two experienced clinicians. First-order, shape, and texture features were extracted using the PyRadiomics library, resulting in 200 radiomics features (186 intensity/texture features and 14 shape features). In addition, rPET (i.e., tumor SUVmax/Liver SUVmax) parameter was included, yielding a grand total of 201 features. The feature set was reduced to 20 variables using ANOVA, with both selection and model evaluation performed via stratified 5-fold cross-validation.

Results: The proposed approach achieved an average balanced accuracy of 0.58 ± 0.07 and an average AUC of 0.71 ± 0.04. Average sensitivity and specificity were 0.48 and 0.68, respectively. The model obtained an average Gmean of 0.57, indicating balanced and stable classification performance. Conclusions: Our ML models trained on PET/CT radiomic features showed moderate discriminatory performance for TET risk stratification.

Computed Tomography (CT) and Positron Emission Tomography (PET) provide complementary anatomical and metabolic information essential for oncology assessment; however, PET imaging remains constrained by radiation exposure, acquisition cost, and limited accessibility. This work proposes a 3D CT-to-PET translation framework that integrates contrastive latent representation learning within a variational autoencoder (VAE) and Brownian Bridge diffusion to effectively bridge structural and functional modality differences. Evaluated on 1,014 paired PET/CT volumes across five anatomical regions-brain, lung, liver, kidney, and stomach-the proposed approach demonstrates improved synthetic PET signal fidelity and reliable lesion representation while maintaining efficient diffusion inference, supporting diagnostically meaningful virtual PET imaging applications.

COVID-19 prognosis using clinical tabular data faces significant challenges due to missing values and class imbalance issues. Existing methods often overlook the complex high-order interrelationship among clinicalattributes and struggle with training stability on imbalanced datasets. We propose ACGM, an attribute-centric graph modeling network that simultaneously addresses missing data imputation and COVID-19 prognosis. ACGM consists of three key modules: an attributes preprocessing module (APM) for coarse-grained imputation initialization, a graph-enhanced attributes imputation module (GEAIM) that models high-order inter-attribute relationships through graph structures, and a graph-enhanced disease prognosis module (GEDPM) that leverages these complex attribute interactions for final prediction. GEAIM and GEDPM employ a mean-teacher strategy with attributes graph matching to preserve high-order relationships, enhance training stability, and maintain structural integrity of attribute interactions. Extensive experiments are conducted on four public COVID-19 tabular datasets, demonstrating the superiority of our ACGM over existing methods. Through comprehensive interpretability analysis, we identify that attributes such as LDH, Difficulty In Breathing, and SaO2 significantly impact COVID-19 prognosis, aligning well with clinical insights and radiologist assessments.

Objective: This study aims to propose a multimodal, multi-view deep learning approach for breast cancer virtual biopsy, a non-invasive classification of breast lesions as malignant or benign, by integrating Full-Field Digital Mammography (FFDM) and Contrast-Enhanced Spectral Mammography (CESM). The work addresses the critical challenge of missing CESM data by introducing generative artificial intelligence (AI) to synthesize CESM images when unavailable, ensuring the continuity of diagnostic workflows.

Methods: The proposed method uses FFDM and CESM images in both craniocaudal (CC) and mediolateral oblique (MLO) views. When CESM is missing, a CycleGAN-based generative model produces synthetic CESM images from FFDM inputs. For classification, three convolutional neural networks (ResNet18, ResNet50, and VGG16) are employed, and a two-stage late fusion strategy integrates view-specific and modality-specific malignancy probabilities, weighted by Matthews Correlation Coefficient (MCC), into a final malignancy score. The system’s robustness under varying degrees of missing CESM data is tested by incrementally replacing real CESM inputs with synthetic ones and evaluating classification performance using AUC, G-mean, and MCC.

Results: CycleGAN achieved high-fidelity CESM synthesis, with Peak-Signal-to-Noise Ratio exceeding 24 dB and Structural Similarity Index above 0.8 across both CC and MLO views. For lesion classification, the multimodal configuration combining FFDM and CESM consistently outperformed the unimodal FFDM-only setup. Notably, even when CESM was entirely replaced by synthetic images, the multimodal approach still improved virtual biopsy performance compared to FFDM alone. Although classification performance declined as the proportion of synthetic CESM increased, the use of synthetic data remained beneficial.

Conclusion: This work demonstrates that generative AI can effectively address missing-modality challenges in breast cancer diagnostics by synthesizing CESM images to enhance FFDM-based virtual biopsy pipelines. In the absence of real CESM data, incorporating synthetic images improves lesion classification compared to using FFDM alone, offering a non-invasive alternative to support clinical decision-making. Moreover, by releasing the extended CESM@UCBM dataset, this study contributes a valuable resource for advancing research and innovation in breast multimodal diagnostic systems.

Background and Objectives: Despite the significant potential of Foundation Models (FMs) in medical imaging, their application to prognosis prediction remains challenging due to data scarcity, class imbalance, and task complexity, limiting their clinical adoption. This study introduces the first structured benchmark to assess the robustness and efficiency of transfer learning strategies for FMs compared with convolutional neural networks (CNNs) in predicting COVID-19 patient outcomes from chest X-rays. The goal is to systematically compare fine-tuning strategies, classical and parameter-efficient, under realistic clinical constraints related to data scarcity and class imbalance, offering empirical guidance for AI deployment in clinical workflows.

Methods: Four publicly available COVID-19 chest X-ray datasets were used, covering mortality, severity, and ICU admission, with varying sample sizes and class imbalances. CNNs pretrained on ImageNet and FMs pretrained on general or biomedical datasets were adapted using full fine-tuning, linear probing, and parameter-efficient methods. Models were evaluated under full-data and few-shot regimes using Matthews Correlation Coefficient (MCC) and Precision–Recall AUC (PR-AUC) with cross-validation and class-weighted losses.

Results: CNNs with full fine-tuning performed robustly on small, imbalanced datasets, while FMs with Parameter-Efficient Fine-Tuning (PEFT), particularly LoRA and BitFit, achieved competitive results on larger datasets. Severe class imbalance degraded PEFT performance, whereas balanced data mitigated this effect. In few-shot settings, FMs showed limited generalization, with linear probing yielding the most stable results.

Conclusions: No single fine-tuning strategy proved universally optimal. CNNs remain dependable for low-resource scenarios, whereas FMs benefit from parameter-efficient methods when data are sufficient.

Background and Objective: The advancement of generative AI in medical imaging faces the trilemma of simultaneously achieving high fidelity and diversity in synthetic data generation. Although Generative Adversarial Networks (GANs) have demonstrated significant potential, they are often hindered by limitations such as mode collapse and poor coverage of real data distributions. This study investigates the use of GAN ensembles as a solution to these challenges, with the goal of enhancing the quality and utility of synthetic medical images.

Methods: We formulate a multi-objective optimisation problem to select an optimal ensemble of GANs that balances fidelity and diversity. The ensemble comprises models that contribute uniquely to the synthetic data space, ensuring minimal redundancy. A comprehensive evaluation was conducted using three distinct medical imaging datasets. We tested 22 GAN architectures, incorporating various loss functions and regularisation techniques. By sampling models at different training epochs, we crafted 110 unique configurations for ensemble selection.

Results: The selected GAN ensembles demonstrated improved performance in generating synthetic medical images that closely resemble real data distributions. These ensembles preserved image fidelity while increasing diversity. In some settings, downstream models trained on synthetic data achieved slightly higher accuracy than those trained on real data alone. This effect arises because the synthetic images act as a targeted data augmentation mechanism that enhances class balance and diversity rather than replacing real data.

Conclusions: GAN ensembles offer a robust solution to the fidelity–diversity–efficiency trade-off in medical image synthesis. By integrating multiple complementary models, the proposed approach improves the representativeness and utility of synthetic medical data, potentially advancing a wide range of clinical and research applications in diagnostic AI.

Data scarcity remains a critical bottleneck impeding technological advancements across various domains, including but not limited to medicine and precision agriculture. To address this challenge, we explore the potential of Deep Generative Models (DGMs) in producing synthetic data that satisfies the Generative Learning Trilemma: fidelity, diversity, and sampling efficiency. However, recognizing that these criteria alone are insufficient for practical applications, we extend the trilemma to include utility, robustness, and privacy, factors crucial for ensuring the applicability of DGMs in real-world scenarios. Evaluating these metrics becomes particularly challenging in data-scarce environments, as DGMs traditionally rely on large datasets to perform optimally. This limitation is especially pronounced in domains like medicine and precision agriculture, where ensuring acceptable model performance under data constraints is vital. To address these challenges, we assess the Generative Learning Trilemma in data-scarcity settings using state-of-the-art evaluation metrics, comparing three prominent DGMs: Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models (DMs). Furthermore, we propose a comprehensive framework to assess utility, robustness, and privacy in synthetic data generated by DGMs. Our findings demonstrate varying strengths among DGMs, with each model exhibiting unique advantages based on the application context. This study broadens the scope of the Generative Learning Trilemma, aligning it with real-world demands and providing actionable guidance for selecting DGMs tailored to specific applications.

Objective: Idiopathic normal pressure hydrocephalus (INPH) is a treatable neurological condition, yet predicting which patients will benefit from a cerebrospinal fluid shunt remains challenging. Structural brain MRI is a core part of the diagnostic workup, but traditional radiological measures show limited predictive accuracy. This study aimed to assess whether deep learning and radiomics-based machine learning approaches can provide clinically useful predictions of shunt outcome based on preoperative MRI.

Methods: We investigated 149 shunted INPH patients with available preoperative T1-weighted, T2-weighted, and FLAIR images. Patients were classified as responders (n = 113) or non-responders (n = 36) based on postoperative gait speed improvement. For INPH, this is a large sample with typical outcome distribution. Three artificial intelligence approaches were tested: a late-fusion ensemble of multiple 3D convolutional neural networks; a multimodal intermediate fusion model; and radiomics-based machine learning models trained on features extracted from whole-brain masks. Models were assessed using 10-fold cross-validation. The best performing model on the validation set was selected from each approach. Performance metrics included the area under the receiving operating characteristic curve (AUROC), sensitivity, and specificity.

Results: Performance was considered poor in all models, and none reached an area under the receiving operating characteristic curve above 70%. Of the three methodologies, the best performance was achieved with a radiomics-based model (Linear Discriminant Analysis classifier on T1-weighted images) which achieved an AUROC of 63.7%. In a reduced subset of clearly separated responders and non-responders (n = 72), the best model (late fusion ensemble of 5 convolutional neural networks) reached an AUROC of 69.2%.

Conclusions: Despite the use of advanced artificial intelligence techniques, structural MRI alone were insufficient for reliably predicting gait outcome after surgery in idiopathic normal pressure hydrocephalus. To capture the complexity of the condition and enable clinically meaningful predictions, our findings indicate the need for research investigating multimodal input and using large multi-center datasets.