This study presents Latent Diffusion Autoencoder (LDAE), a novel encoder-decoder diffusion-based framework for efficient and meaningful unsupervised learning in medical imaging, focusing on Alzheimer’s disease (AD) using brain MRI from the ADNI database as a case study. Unlike conventional diffusion autoencoders operating in image space, LDAE applies the diffusion process in a compressed latent representation, improving computational efficiency and making 3D medical imaging representation learning tractable. To validate the proposed approach, we explore two key hypotheses: (i) LDAE effectively captures meaningful semantic representations on 3D brain MRI associated with AD and ageing, and (ii) LDAE achieves high-quality image generation and reconstruction while being computationally efficient. Experimental results support both hypotheses: (i) linear-probe evaluations demonstrate promising diagnostic performance for AD (AUROC: 90%, ACC: 84%) and age prediction (MAE: 4.1 years, RMSE: 5.2 years); (ii) the learned semantic representations enable attribute manipulation, yielding anatomically plausible modifications; (iii) semantic interpolation experiments show strong reconstruction of missing scans, with SSIM of 0.969 (MSE: 0.0019) for a 6-month gap. Even for longer gaps (24 months), the model maintains robust performance (SSIM  >  0.93, MSE  <  0.004), indicating an ability to capture temporal progression trends; (iv) compared to conventional diffusion autoencoders, LDAE significantly increases inference throughput (20 ×  faster) while also enhancing reconstruction quality. These findings position LDAE as a promising framework for scalable medical imaging applications, with the potential to serve as a foundation model for medical image analysis. Code is publicly available at https://github.com/GabrieleLozupone/LDAE.

Accurate overall survival (OS) prediction in non-small cell lung cancer (NSCLC) is crucial but challenging due to high-dimensional 3D computed tomography (CT) data, limited annotations, and time-to-event outcomes. Traditional 3D CNNs are computationally expensive and prone to overfitting on small datasets. We propose a lightweight framework that aggregates 2D CT slice embeddings via soft attention to form a 3D patient representation. In our approach, features are extracted with EfficientNetB0, and DeepHit models time-to-event survival. Validated on LUNG1 (415 patients), our method outperforms 3D ResNet (+0.077) and alternative aggregation strategies (+0.005) in time-dependent concordance index (Ctd-index). Furthermore, transfer learning from LUNG1 improves performance on the small private CLARO dataset (0.579 vs 0.503). This shows that 2D CNNs with soft attention provide a computationally efficient yet effective alternative to 3D CNN architectures for NSCLC OS prediction, with a substantially lower computational cost (54.3 GFLOPs vs. 2924.6 GFLOPs for ResNet3D-18).

Digital Breast Tomosynthesis (DBT) is a key imaging modality for breast cancer detection, improving lesion visibility by reducing tissue overlap inherent in conventional mammography. In this work, we propose a novel deep learning framework that classifies DBT volumes as malignant or non-malignant, while simultaneously generating a synthetic 2D image to assist diagnostic interpretation. This image is derived from a 3D saliency map computed by the internal attention mechanisms of the model, which highlights and preserves the most diagnostically relevant regions from the original volume. A surface is defined in this saliency space, enabling sampling and projection into a 2D diagnostic representation. This projection offers a compact summary of the volumetric scan, assisting clinicians in diagnostic interpretation and potentially alleviating the cognitive workload. A standard convolutional neural network trained on these synthetic 2D images achieves classification performance comparable to models operating directly on full 3D volumes. We train and evaluate our method on the OPTI-MAM dataset and assess generalization through external validation on the independent BCS-DBT dataset without retraining. Results show that the model performs robustly across different clinical sources and provides an interpretable, computationally efficient tool for DBT-based breast cancer diagnosis.

Stroke is one of the leading causes of death and long-term disability worldwide, with large vessel occlusion representing one of the most severe forms due to its association with extensive brain damage and poor prognosis. Rapid and reliable detection of large vessel occlusion in emergency settings is therefore essential to guide timely treatment decisions. Computed tomography angiography is currently the reference imaging modality for this task, as it provides high-resolution visualization of cerebral vessels within minutes. Nevertheless, the small size and variable location of thrombi make their identification difficult, often requiring expert radiological interpretation and being prone to missed detections. In this study, we propose a unified deep learning architecture that integrates GravityNet for slice-wise localization of vessel obstruction with a multi-head self-attention mechanism designed to capture spatial continuity across adjacent slices. A volumetric refinement stage based on three-dimensional non-maximum suppression consolidates overlapping predictions and reduces false positives across the brain volume. Evaluated on a private dataset of computed tomography angiography scans, the proposed method achieves 70.8% sensitivity at one false positive per scan, showing its potential to support automated and time-critical detection of clots in acute stroke workflows.

This work presents an advanced machine-learning framework for high-resolution urban air-quality forecasting, developed within the PROAIR project to power environmental Digital Twin simulations. We propose a hybrid pipeline that integrates regional forecasts from the Copernicus Atmosphere Monitoring Service (CAMS) with localized ground-truth data from EEA/ARPA monitoring networks. To overcome the spatial coarseness of global models, we implement a data augmentation layer using a local distance-based spatial augmentation and temporal gap-filling to generate a continuous, high-resolution urban dataset. The core predictive engine leverages a Temporal Fusion Transformer (TFT) architecture, specifically optimized with a multi-head attention mechanism to capture long-range temporal dependencies and a custom weighted loss function to prioritize the detection of critical pollution peaks. Our framework supports interactive "What-if" scenario analysis, allowing stakeholders to simulate the impact of environmental and anthropogenic modifiers—such as traffic restrictions or weather shifts—on local pollutant concentrations. Results demonstrate that the TFT model significantly outperforms baseline regressions in stability and peak-event accuracy. The system’s modular design ensures it is computationally efficient for real-time integration into smart-city decision-making platforms and future edge-AI deployments.

Accurate prognosis of Non-Small Cell Lung Cancer (NSCLC) is crucial for enhancing patient care and treatment outcomes. Despite the advancements in deep learning, the task of overall survival prediction in NSCLC has not fully leveraged these techniques, yet. This study introduces a novel methodology for predicting 2-year overall survival (OS) in NSCLC patients using CT scans. Our approach integrates CT scan representations produced by EfficientNetB0 with a soft attention mechanism to identify the most relevant slices for survival risk prediction, which are then analyzed by a risk-assessment network. To validate our method and ensure reproducibility, we employed the public LUNG1 dataset and a smaller private dataset. Our approach was compared to benchmark 3D networks and two variants of our methodology: on the LUNG1 it outperformed the competitors achieving a mean C^td -index of 0.584 over tenfold cross-validation. On the LUNG1 we also demonstrated the adaptability of our method with 5 other 2D backbones replacing the EfficientNetB0, confirming that our mechanism of combining 2D slice representations to construct a 3D volume representation is more effective for OS prediction compared to a traditional 3D approach. Finally, we used transfer learning on the private dataset, showing that it can significantly enhance performance in limited data scenarios, increasing the C^td -index by 0.076 compared to model without transfer learning.

Background: Large vessel occlusion (LVO) is a major cause of severe ischemic stroke and requires rapid diagnosis to enable timely treatment. Computed tomography angiography (CTA) is widely used to visualize cerebral vessels and detect occlusions, but manual interpretation of CTA volumes is challenging and time-consuming. Existing automated approaches typically analyze individual slices or single-view projections and therefore do not fully exploit the complementary information available across different anatomical views. Methods: We propose a cross-view detection approach for automated LVO identification in CTA volumes. A slice-based detector is applied independently to axial, coronal, and sagittal views to generate candidate detections. These detections are mapped into a common volumetric coordinate system and analyzed jointly. Spatial neighborhoods are identified using a KD-tree, and candidate detections are retained only if supported by multiple anatomical views. The remaining candidates are refined using three-dimensional non-maximum suppression. Results: The method was evaluated on an in-house dataset of 130 CTA scans with expert LVO annotations. The proposed approach achieved a sensitivity of 79.2% at one false positive per case and an area under the FROC curve of 87%. Conclusions: Exploiting cross-view spatial consistency improves automated LVO detection in CTA volumes and provides a robust strategy for integrating complementary information across anatomical views.

Accurate survival prediction in non-small cell lung cancer (NSCLC) requires integrating clinical, radiological, and histopathological data. Multimodal deep learning (MDL) can improve precision prognosis, but small cohorts and missing modalities limit its clinical applicability, as conventional approaches enforce complete-case filtering or imputation. We present a missing-aware multimodal survival framework that combines computed tomography (CT), whole-slide histopathology images (WSI), and structured clinical variables for overall survival modeling in unresectable stage II-III NSCLC. The framework uses foundation models (FMs) for modality-specific feature extraction and a missing-aware encoding strategy that enables intermediate multimodal fusion under naturally incomplete modality profiles. By design, the architecture processes all available data without dropping patients during training or inference. Intermediate fusion outperforms unimodal baselines and both early and late fusion strategies, with the trimodal configuration reaching a C-index of 74.42. Modality-importance analyses show that the fusion model adapts its reliance on each data stream according to representation informativeness, shaped by the alignment between FM pretraining objectives and the survival task. The learned risk scores produce clinically meaningful stratification of disease progression and metastatic risk, with statistically significant log-rank tests across all modality combinations, supporting the translational relevance of the proposed framework.

Pulmonary embolism (PE) is a life-threatening condition for which accurate risk stratification is essential for clinical decision-making. Most existing prognostic models rely on a single data modality, typically structured electronic health records (EHR), limiting their ability to capture the heterogeneous nature of the disease. Multimodal approaches combining EHR, computed tomography (CT) imaging, and free-text clinical reports provide complementary information, but the optimal fusion strategy for prognosis prediction remains unclear. In this study, we present a comprehensive comparative study of multimodal fusion strategies for PE prognosis using the INSPECT dataset, which includes structured EHR, CT images, and clinical text with mortality labels. We develop unimodal baselines and multiple bimodal and trimodal models using early, intermediate, late, and cross-fusion strategies within a unified framework. Models are evaluated for 1-, 6-, and 12-month mortality prediction. Multimodal models consistently outperform unimodal baselines across all prediction horizons, with MCC improvements up to +0.13 for 1-month mortality (0.399 vs 0.269) and +0.11 for 6-month mortality (0.479 vs 0.367). The best overall performance is achieved by late fusion models, reaching an MCC of 0.497 for 12-month prediction, compared to 0.454 for the best unimodal model. Interestingly, bimodal combinations often outperform trimodal fusion in short-term prediction, suggesting that improper integration of multiple modalities may introduce noise and reduce effectiveness. These results highlight the importance of fusion design in multimodal clinical AI and provide practical guidance for developing robust prognosis models for pulmonary embolism. Code is publicly available at https://github.com/arco-group/INSPECT-CS.

Purpose To investigate whether selective removal of vascular structures can improve lesion visibility and interpretability in maximum intensity projection (MIP) images derived from dynamic contrast-enhanced MRI. Materials and Methods A retrospective analysis was conducted using breast MRI scans from the Duke-Breast-Cancer-MRI (Duke) dataset for model development. DeepVEST, a deep learning method for automatic vessel segmentation and removal, was developed. A reader study with five breast radiologists was conducted to evaluate the impact of vessel-removed MIPs on lesion assessment, using images from the Duke and Advanced-MRI-Breast-Lesions (AMBL) datasets. Segmentation performance was evaluated against manually annotated vessel segmentations using the Dice similarity coefficient (DSC), and perceived usefulness and quality were measured through reader study metrics. Interreader agreement on vessel removal quality and artifact impact was calculated using the Gwet agreement coefficient (AC1). Results DeepVEST achieved a DSC of 0.611 for vessel segmentation. In the reader study (150 assessments, 600 responses), vessels partially or fully obscured lesion margins in 91 of 150 assessments (60.7%). Vessel removal effectiveness was rated 3.820 ± 0.749 on a 5-point Likert scale. Artifacts were reported in 45 of 150 assessments (30.0%), with a low average severity score of 0.493 ± 0.900 (scale 0-5), indicating minimal image quality impact. Substantial interreader agreement was observed for vessel removal quality (AC1 = 0.703) and artifact evaluation (AC1 = 0.736). Conclusion DeepVEST enabled automatic vessel segmentation and removal in breast MRI MIP images, showing high effectiveness with minimal artifacts. Most readers found it helpful for lesion assessment in challenging cases. ©RSNA, 2026.

Large Vessel Occlusion is one of the most critical neurological emergencies in stroke care, requiring rapid and accurate diagnosis to optimize clinical outcomes. Automated detection tools have demonstrated the potential to significantly reduce treatment time, thereby improving patient prognosis. In this paper, we propose a novel two-stage deep learning approach for detecting large vessel occlusion and assessing its volume directly from computed tomography angiography. The first stage employs a two-dimensional convolutional neural network-based detector, built upon GravityNet, specifically adapted for single lesion detection with a novel pixel-based configuration. The second stage applies a three-dimensional false-positive reduction technique to refine predictions within the brain volume. Our method achieves 80% sensitivity at two false positives per case, demonstrating its robustness and effectiveness in detecting large vessel occlusions on computed tomography angiography.

Nuclei detection in cervical cytology images holds significant promise for enhancing the early identification of cervical cancer. In this study, we apply GravityNet, a one-stage end-to-end detector designed ad hoc for detecting small regions, like cervical cell nuclei. This approach is based on pixel-based anchors, which serve as reference points on the image pixel grid and dynamically move towards the suspicious regions by inspecting the surrounding space. We conducted experiments on the publicly available Cervix93 dataset to evaluate the effectiveness of our method. Despite challenges posed by variations in staining techniques and cell morphology, our results show promising performance compared to the current state-of-the-art methods.

This study presents Latent Diffusion Autoencoder (LDAE), a novel encoder-decoder diffusion-based framework for efficient and meaningful unsupervised learning in medical imaging, focusing on Alzheimer disease (AD) using brain MR from the ADNI database as a case study. Unlike conventional diffusion autoencoders operating in image space, LDAE applies the diffusion process in a compressed latent representation, improving computational efficiency and making 3D medical imaging representation learning tractable. To validate the proposed approach, we explore two key hypotheses: (i) LDAE effectively captures meaningful semantic representations on 3D brain MR associated with AD and ageing, and (ii) LDAE achieves high-quality image generation and reconstruction while being computationally efficient. Experimental results support both hypotheses: (i) linear-probe evaluations demonstrate promising diagnostic performance for AD (ROC-AUC: 90

Battery State-of-Health (SoH) estimation is essential for ensuring the reliability, safety, and cost-effectiveness of lithium-ion batteries in electric vehicles. Traditional approaches often require detailed knowledge of battery degradation mechanisms, making them computationally expensive and difficult to generalize. This paper introduces a data-driven methodology focused on lightweight features extracted exclusively from the final portion of the constant current (CC) charging phase, commonly observed in real-world scenarios. Statistical and shape-based features, such as mean, standard deviation, skewness, kurtosis, charging duration, accumulated charge, slope, and entropy, were computed from voltage and current signals. Several machine learning models, including Support Vector Regressors, Random Forest, Linear Regression, K-Nearest Neighbors, and Multi-Layer Perceptrons, were evaluated for accuracy, computational complexity, and suitability for embedded deployment. Extensive experimentation on a dataset comprising 92 lithium-ion cells demonstrates that the proposed approach achieves robust predictive performance, with mean absolute percentage errors around 1

The automated processing of Electronic Health Records (EHRs) poses a significant challenge due to their unstructured nature, rich in valuable, yet disorganized information. Natural Language Processing (NLP), particularly Named Entity Recognition (NER), has been instrumental in extracting structured information from EHR data. However, existing literature primarly focuses on extracting handcrafted clinical features through NLP and NER methods without delving into their learned representations. In this work, we explore the untapped potential of these representations by considering their contextual richness and entity-specific information. Our proposed methodology extracts representations generated by a transformer-based NER model on EHRs data, combines them using a hierarchical attention mechanism, and employs the obtained enriched representation as input for a clinical prediction model. Specifically, this study addresses Overall Survival (OS) in Non-Small Cell Lung Cancer (NSCLC) using unstructured EHRs data collected from an Italian clinical centre encompassing 838 records from 231 lung cancer patients. Whilst our study is applied on EHRs written in Italian, it serves as use case to prove the effectiveness of extracting and employing high level textual representations that capture relevant information as named entities. Our methodology is interpretable because the hierarchical attention mechanism highlights the information in EHRs that the model considers the most crucial during the decision-making process. We validated this interpretability by measuring the agreement of domain experts on the importance assigned by the hierarchical attention mechanism to EHRs information through a questionnaire. Results demonstrate the effectiveness of our method, showcasing statistically significant improvements over traditional manually extracted clinical features.

Predicting pathologic complete response in non-small cell lung cancer is crucial for tailoring effective treatment strategies and to improve patient outcomes. With the increasing application of artificial intelligence in cancer research, machine learning is poised to play a significant role in prognostication and decision-making. This paper presents a novel approach that utilizes named entity recognition and attention mechanisms applied to electronic health records to predict the pathologic complete response. We first employ named entity recognition to extract relevant biomedical entities from unstructured clinical notes within reports. These entities, combined with structured data, are then processed using a hierarchical attention mechanism to generate comprehensive patient representations. This approach captures complex relationships and contextual information within electronic health records compared to traditional methods. The results highlight the potential of advanced natural language processing techniques to enhance clinical decision-making and support personalized treatment planning in oncology.

Covid-19 infection influenced the screening test rate of breast cancer worldwide due to the quarantine measures, routine procedures reduction, and delay of early diagnosis, causing high mortality risk and severity of the disease. X-ray mammography is the gold standard for diagnosing early signs of breast cancer, and Artificial Intelligence enables the detection of suspicious lesions and classifying them in terms of malignancy. This paper aimed to investigate mass detection and classification in a large-scale OPTIMAM dataset with 6000 cases and extracted 3524 images with masses in the mammograms of the Hologic manufacturer. The methodology of the detection step is to train the RetinaNet architecture of ResNet50, ResNet101, and ResNet152 backbones with three types of initializations by ImageNet and COCO weights and from scratch. The dataset was pre-processed to generate two types of input with entire mammograms and patches, which are stated as the first and the second approaches. The results show that in the first approach, RetinaNet of ResNet50 backbone with ImageNet and COCO weights and ResNet152 with the same weights performed 0.91 True Positive Rate at 0.78 False Positive Per Image, respectively. In contrast, in the second approach, ResNet152 with ImageNet weights reached 0.88 TPR at 0.78 FPPI. In the classification step, the Transfer Learning approach was applied with fine-tuning by adding L2-regularization and class weights to balance class distribution in the datasets.