The paper’s main contributions are twofold: to demonstrate how to apply the general European Union’s High-Level Expert Group’s (EU HLEG) guidelines for trustworthy AI in practice for the domain of healthcare; and to investigate the research question of what does “trustworthy AI” mean at the time of the COVID-19 pandemic. To this end, we present the results of a post-hoc self-assessment to evaluate the trustworthiness of an AI system for predicting a multi-regional score conveying the degree of lung compromise in COVID-19 patients, developed and verified by an interdisciplinary team with members from academia, public hospitals, and industry in time of pandemic. The AI system aims to help radiologists to estimate and communicate the severity of damage in a patient’s lung from Chest X-rays. It has been experimentally deployed in the radiology department of the ASST Spedali Civili clinic in Brescia (Italy) since December 2020 during pandemic time. The methodology we have applied for our post-hoc assessment, called Z-Inspection, uses socio-technical scenarios to identify ethical, technical and domain-specific issues in the use of the AI system in the context of the pandemic.

Purpose: This study aims at exploiting artificial intelligence (AI) for the identification, segmentation and quantification of COVID-19 pulmonary lesions. The limited data availability and the annotation quality are relevant factors in training AI-methods. We investigated the effects of using multiple datasets, heterogeneously populated and annotated according to different criteria. Methods: We developed an automated analysis pipeline, the LungQuant system, based on a cascade of two U-nets. The first one (U-net $$_1$$1 ) is devoted to the identification of the lung parenchyma; the second one (U-net $$_2$$2 ) acts on a bounding box enclosing the segmented lungs to identify the areas affected by COVID-19 lesions. Different public datasets were used to train the U-nets and to evaluate their segmentation performances, which have been quantified in terms of the Dice Similarity Coefficients. The accuracy in predicting the CT-Severity Score (CT-SS) of the LungQuant system has been also evaluated. Results: Both the volumetric DSC (vDSC) and the accuracy showed a dependency on the annotation quality of the released data samples. On an independent dataset (COVID-19-CT-Seg), both the vDSC and the surface DSC (sDSC) were measured between the masks predicted by LungQuant system and the reference ones. The vDSC (sDSC) values of 0.95±0.01 and 0.66±0.13 (0.95±0.02 and 0.76±0.18, with 5 mm tolerance) were obtained for the segmentation of lungs and COVID-19 lesions, respectively. The system achieved an accuracy of 90% in CT-SS identification on this benchmark dataset. Conclusion: We analysed the impact of using data samples with different annotation criteria in training an AI-based quantification system for pulmonary involvement in COVID-19 pneumonia. In terms of vDSC measures, the U-net segmentation strongly depends on the quality of the lesion annotations. Nevertheless, the CT-SS can be accurately predicted on independent test sets, demonstrating the satisfactory generalization ability of the LungQuant.

We propose and evaluate a procedure for the explainability of a breast density deep learning based classifier. A total of 1662 mammography exams labeled according to the BI-RADS categories of breast density was used. We built a residual Convolutional Neural Network, trained it and studied the responses of the model to input changes, such as different distributions of class labels in training and test sets and suitable image pre-processing. The aim was to identify the steps of the analysis with a relevant impact on the classifier performance and on the model explainability. We used the grad-CAM algorithm for CNN to produce saliency maps and computed the Spearman’s rank correlation between input images and saliency maps as a measure of explanation accuracy. We found that pre-processing is critical not only for accuracy, precision and recall of a model but also to have a reasonable explanation of the model itself. Our CNN reaches good performances compared to the state-of-art and it considers the dense pattern to make the classification. Saliency maps strongly correlate with the dense pattern. This work is a starting point towards the implementation of a standard framework to evaluate both CNN performances and the explainability of their predictions in medical image classification problems.

Computed tomography (CT) scans are used to evaluate the severity of lung involvement in patients affected by COVID-19 pneumonia. Here, we present an improved version of the LungQuant automatic segmentation software (LungQuantv2), which implements a cascade of three deep neural networks (DNNs) to segment the lungs and the lung lesions associated with COVID-19 pneumonia. The first network (BB-net) defines a bounding box enclosing the lungs, the second one (U-net $$_1$$1 ) outputs the mask of the lungs, and the final one (U-net $$_2$$2 ) generates the mask of the COVID-19 lesions. With respect to the previous version (LungQuantv1), three main improvements are introduced: the BB-net, a new term in the loss function in the U-net for lesion segmentation and a post-processing procedure to separate the right and left lungs. The three DNNs were optimized, trained and tested on publicly available CT scans. We evaluated the system segmentation capability on an independent test set consisting of ten fully annotated CT scans, the COVID-19-CT-Seg benchmark dataset. The test performances are reported by means of the volumetric dice similarity coefficient (vDSC) and the surface dice similarity coefficient (sDSC) between the reference and the segmented objects. LungQuantv2 achieves a vDSC (sDSC) equal to 0.96 ± 0.01 (0.97 ± 0.01) and 0.69 ± 0.08 (0.83 ± 0.07) for the lung and lesion segmentations, respectively. The output of the segmentation software was then used to assess the percentage of infected lungs, obtaining a Mean Absolute Error (MAE) equal to 2%.

Artificial intelligence (AI) is widely expected to transform healthcare, yet its adoption in clinical practice remains limited. This paper examines the perspectives of Italian clinicians and medical physicists on the drivers of and barriers to AI use. Using an online survey of healthcare professionals across different domains, we find that efficiency gains—such as reducing repetitive tasks and accelerating diagnostics—are the strongest incentives for adoption. However, trust in AI systems, explainability, and the limited availability of AI tools are major obstacles. Respondents emphasized that AI should augment, not replace, medical expertise, calling for participatory development processes where clinicians are actively involved in the design and validation of decision support tools. At the organizational level, the adoption of AI tools is facilitated by innovation-oriented leadership and sufficient resources, while conservative management and economic constraints hinder implementation. The awareness of regulatory frameworks, including the EU AI Act, is moderate, and many clinicians express the need for targeted training to support safe integration. Our findings suggest that the successful adoption of AI in healthcare will depend on building trust through transparency, clarifying legal responsibilities, and fostering organizational cultures that support collaboration between humans and AI. The role of AI in medicine is therefore best understood as a complement to clinical judgment, rather than a replacement.