Alessandra Marasi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Alessandra Marasi has a total of 67 co-authors. Here are the top co-authors by number of shared publications:
- Alessandro De Benedictis
Neurosurgery Unit Bambino Gesù Children's Hospital, IRCCS Rome Italy
4 shared publications - Antonio Napolitano
Medical Physics Department, Bambino Gesù Children's Hospital, IRCCS, Rome, Italy
4 shared publications - Carlo Efisio Marras
Neurosurgery Unit, Bambino Gesù Children’s Hospital, IRCCS, 00165 Rome, Italy
4 shared publications Alberto Redaelli
2 shared publications- Andrea Carai
Neurosurgery Unit Bambino Gesù Children's Hospital, IRCCS Rome Italy
2 shared publications - Antonella Bua
Neurosurgery Unit, Bambino Gesù Children’s Hospital, IRCCS, 4, Piazza S. Onofrio, 00165 Rome, Italy
2 shared publications
Latest Publications
- Open Access
- 1 October 2026
- Published by Elsevier in SSRN Electronic Journal
Benchmarking and error profiling of survival prediction models in glioblastoma MRI
Background and Objective: Deep learning (DL) is increasingly used to predict overall survival (OS) in glioblastoma (GB), yet it rarely outperforms conventional machine learning (ML), and models are compared only through aggregate metrics that do not show which patients they mispredict.Methods: We evaluated a clinical and a radiomic Cox model and four pretrained DL architectures (two brain-specific, two general medical) on an institutional cohort (n=297) and the public UCSF-PDGM cohort (n=350). Discrimination was measured with the concordance index (C-index) under within-cohort cross-validation, cross-cohort comparison of model rankings, and bidirectional external validation. We further introduced a censoring-aware error profiling procedure based on Martingale residuals, which extends to survival prediction the analysis of hidden failure subgroups developed for classification, and characterized mispredicted patients in the model embeddings and by clinical and tumor characteristics.Results: The C-index was 0.60–0.68 for all models; ML performed as well as DL, and general medical pretraining matched brain-specific pretraining. Under external validation, the C-index fell to 0.56–0.63 in both directions. DL models agreed with each other no more than with the ML baselines on which patients they mispredicted, and the error groups were not separable inthe embedding space. In the institutional cohort, underestimated-risk patients had larger T2-FLAIR tumors and were more often MGMT-unmethylated (q < 0.05); this profile was not found in the public cohort. Saliency maps were spatially diffuse, with tumor activation only 1.2 times that of the surrounding brain.Conclusions: The proposed framework compares survival models, quantifies how much DL adds over traditional ML, and adds transparency about which patients are mispredicted and which characteristics are associated with the error. The procedure is model-agnostic, applies to any survival model, and makes model comparison informative for individual patients rather than only on average.
- 1 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBenchmarking and error profiling of survival prediction models in glioblastoma MRI
Background and Objective: Deep learning (DL) is increasingly used to predict overall survival (OS) in glioblastoma (GB), yet it rarely outperforms conventional machine learning (ML), and models are compared only through aggregate metrics that do not show which patients they mispredict.Methods: We evaluated a clinical and a radiomic Cox model and four pretrained DL architectures (two brain-specific, two general medical) on an institutional cohort (n=297) and the public UCSF-PDGM cohort (n=350). Discrimination was measured with the concordance index (C-index) under within-cohort cross-validation, cross-cohort comparison of model rankings, and bidirectional external validation. We further introduced a censoring-aware error profiling procedure based on Martingale residuals, which extends to survival prediction the analysis of hidden failure subgroups developed for classification, and characterized mispredicted patients in the model embeddings and by clinical and tumor characteristics.Results: The C-index was 0.60–0.68 for all models; ML performed as well as DL, and general medical pretraining matched brain-specific pretraining. Under external validation, the C-index fell to 0.56–0.63 in both directions. DL models agreed with each other no more than with the ML baselines on which patients they mispredicted, and the error groups were not separable inthe embedding space. In the institutional cohort, underestimated-risk patients had larger T2-FLAIR tumors and were more often MGMT-unmethylated (q < 0.05); this profile was not found in the public cohort. Saliency maps were spatially diffuse, with tumor activation only 1.2 times that of the surrounding brain.Conclusions: The proposed framework compares survival models, quantifies how much DL adds over traditional ML, and adds transparency about which patients are mispredicted and which characteristics are associated with the error. The procedure is model-agnostic, applies to any survival model, and makes model comparison informative for individual patients rather than only on average.
- Open Access
- 4 December 2025
- Published by Springer Nature in International Journal of Computer Assisted Radiology and Surgery
- Vol. 21 (3) , 473-482
Glioblastoma survival prediction through MRI and clinical data integration with transfer learning
Purpose: Accurate prediction of overall survival (OS) in glioblastoma patients is critical for advancing personalized treatments and improving clinical trial design. Conventional radiomics approaches rely on manually engineered features, which limit their ability to capture complex, high-dimensional imaging patterns. This study employs a deep learning architecture to process MRI data for automated glioma segmentation and feature extraction, leveraging high-level representations from the encoder’s latent space. Methods: Multimodal MRI data from the BraTS2020 dataset and a proprietary dataset from Fondazione IRCCS Istituto Neurologico Carlo Besta (Milan, Italy) were processed independently using a U-Net-like model pre-trained on BraTS2018 and fine-tuned on BraTS2020. Features extracted from the encoder’s latent space represented hierarchical imaging patterns. These features were combined with clinical variable (patient’s age) and reduced via principal component analysis (PCA) to enhance computational efficiency. Machine learning classifiers—including random forest, XGBoost, and a fully connected neural network—were trained on the reduced feature vectors for OS classification. Results: In the four-modality BraTS4CH setting, the multi-layer perceptron achieved the best performance (F1 = 0.71, AUC = 0.74, accuracy = 0.71). When limited to two modalities on BraTS2020 (BraTS2CH), MLP again led (F1 = 0.67, AUC = 0.70, accuracy = 0.67). On the IRCCS Besta two-modality cohort (Besta2CH), XGBoost produced the highest F1-score and accuracy (F1 = 0.65, accuracy = 0.66), while MLP obtained the top AUC (0.70). These results are competitive with—and in some metrics exceed—state-of-the-art reports, demonstrating the robustness and scalability of our automated framework relative to traditional radiomics and AI-driven approaches. Conclusion: Integrating encoder-derived features from multimodal MRI data with clinical variables offers a scalable and effective approach for OS prediction in glioblastoma patients. This study demonstrates the potential of deep learning to address traditional radiomics limitations, paving the way for more precise and personalized prognostic tools.
- 4 December 2025
- Published by Springer Nature in International Journal of Computer Assisted Radiology and Surgery
- Vol. 21 (3) , 473-482
Open AccessGlioblastoma survival prediction through MRI and clinical data integration with transfer learning
Purpose: Accurate prediction of overall survival (OS) in glioblastoma patients is critical for advancing personalized treatments and improving clinical trial design. Conventional radiomics approaches rely on manually engineered features, which limit their ability to capture complex, high-dimensional imaging patterns. This study employs a deep learning architecture to process MRI data for automated glioma segmentation and feature extraction, leveraging high-level representations from the encoder’s latent space. Methods: Multimodal MRI data from the BraTS2020 dataset and a proprietary dataset from Fondazione IRCCS Istituto Neurologico Carlo Besta (Milan, Italy) were processed independently using a U-Net-like model pre-trained on BraTS2018 and fine-tuned on BraTS2020. Features extracted from the encoder’s latent space represented hierarchical imaging patterns. These features were combined with clinical variable (patient’s age) and reduced via principal component analysis (PCA) to enhance computational efficiency. Machine learning classifiers—including random forest, XGBoost, and a fully connected neural network—were trained on the reduced feature vectors for OS classification. Results: In the four-modality BraTS4CH setting, the multi-layer perceptron achieved the best performance (F1 = 0.71, AUC = 0.74, accuracy = 0.71). When limited to two modalities on BraTS2020 (BraTS2CH), MLP again led (F1 = 0.67, AUC = 0.70, accuracy = 0.67). On the IRCCS Besta two-modality cohort (Besta2CH), XGBoost produced the highest F1-score and accuracy (F1 = 0.65, accuracy = 0.66), while MLP obtained the top AUC (0.70). These results are competitive with—and in some metrics exceed—state-of-the-art reports, demonstrating the robustness and scalability of our automated framework relative to traditional radiomics and AI-driven approaches. Conclusion: Integrating encoder-derived features from multimodal MRI data with clinical variables offers a scalable and effective approach for OS prediction in glioblastoma patients. This study demonstrates the potential of deep learning to address traditional radiomics limitations, paving the way for more precise and personalized prognostic tools.
- Open Access
Simulators with Haptic Feedback in Neurosurgery: Are We Reaching the “Aviator” Type of Training? Narrative Review and Future Perspectives
Over the last decade, the quality of neurosurgical procedures dramatically improved, also thanks to the development and increased accessibility of several technological recourses (e.g., imaging, neuronavigation, neurophysiology, microscopy), allowing to plan increasingly complex approaches, while reducing the risk of postoperative complications. Among these resources, three-dimensional rendering and simulation systems, such as virtual and augmented reality, provide a high-quality visual reconstruction of brain structures and interaction with advanced anatomical models. Although the usefulness of these systems is now widely recognized, the additional availability of proprioceptive (haptic) feedback might help to further enhance the realism of surgical simulation. A systematic literature review on the application of haptic technology in simulation of cranial neurosurgical procedures was made. Inclusion criteria were the usage of simulators with haptic feedback for specific neurosurgical procedures whereas the studies that did not include an evaluation of the surgical simulation system by a surgeon were excluded. According to inclusion and exclusion criteria, 10 studies were selected. Simulation in neurosurgery still lacks a system capable of rehearsing the entire procedure—from skin incision to skin closure—while providing both visual and proprioceptive feedback. Consequently, further advancements in this area are necessary.
Open AccessSimulators with Haptic Feedback in Neurosurgery: Are We Reaching the “Aviator” Type of Training? Narrative Review and Future Perspectives
Over the last decade, the quality of neurosurgical procedures dramatically improved, also thanks to the development and increased accessibility of several technological recourses (e.g., imaging, neuronavigation, neurophysiology, microscopy), allowing to plan increasingly complex approaches, while reducing the risk of postoperative complications. Among these resources, three-dimensional rendering and simulation systems, such as virtual and augmented reality, provide a high-quality visual reconstruction of brain structures and interaction with advanced anatomical models. Although the usefulness of these systems is now widely recognized, the additional availability of proprioceptive (haptic) feedback might help to further enhance the realism of surgical simulation. A systematic literature review on the application of haptic technology in simulation of cranial neurosurgical procedures was made. Inclusion criteria were the usage of simulators with haptic feedback for specific neurosurgical procedures whereas the studies that did not include an evaluation of the surgical simulation system by a surgeon were excluded. According to inclusion and exclusion criteria, 10 studies were selected. Simulation in neurosurgery still lacks a system capable of rehearsing the entire procedure—from skin incision to skin closure—while providing both visual and proprioceptive feedback. Consequently, further advancements in this area are necessary.
- 1 June 2024
- Published by Edizioni Minerva Medica in Journal of Neurosurgical Sciences
- Vol. 68 (4)
Successful staged separation of total angular craniopagus
Article: Successful staged separation of total angular craniopagus - Journal of Neurosurgical Sciences 2022 Mar 17 - Minerva Medica - Journals
- 1 June 2024
- Published by Edizioni Minerva Medica in Journal of Neurosurgical Sciences
- Vol. 68 (4)
Successful staged separation of total angular craniopagus
Article: Successful staged separation of total angular craniopagus - Journal of Neurosurgical Sciences 2022 Mar 17 - Minerva Medica - Journals
- Open Access
- 31 May 2023
- Published by MDPI AG in Journal of Clinical Medicine
- Vol. 12 (11) , 3779
Vertical Hemispherotomy: Contribution of Advanced Three-Dimensional Modeling for Presurgical Planning and Training
Vertical hemispherotomy is an effective treatment for many drug-resistant encephalopathies with unilateral involvement. One of the main factors influencing positive surgical results and long-term seizure freedom is the quality of disconnection. For this reason, perfect anatomical awareness is mandatory during each step of the procedure. Although previous groups attempted to reproduce the surgical anatomy through schematic representations, cadaveric dissections, and intraoperative photographs and videos, a comprehensive understanding of the approach may still be difficult, especially for less experienced neurosurgeons. In this work, we reported the application of advanced technology for three-dimensional (3D) modeling and visualization of the main neurova-scular structures during vertical hemispherotomy procedures. In the first part of the study, we built a detailed 3D model of the main structures and landmarks involved during each disconnection phase. In the second part, we discussed the adjunctive value of augmented reality systems for the management of the most challenging etiologies, such as hemimegalencephaly and post-ischemic encephalopathy. We demonstrated the contribution of advanced 3D modeling and visualization to enhance the quality of anatomical representation and interaction between the operator and model according to a surgical perspective, optimizing the quality of presurgical planning, intraoperative orientation, and educational training.
- 31 May 2023
- Published by MDPI AG in Journal of Clinical Medicine
- Vol. 12 (11) , 3779
Open AccessVertical Hemispherotomy: Contribution of Advanced Three-Dimensional Modeling for Presurgical Planning and Training
Vertical hemispherotomy is an effective treatment for many drug-resistant encephalopathies with unilateral involvement. One of the main factors influencing positive surgical results and long-term seizure freedom is the quality of disconnection. For this reason, perfect anatomical awareness is mandatory during each step of the procedure. Although previous groups attempted to reproduce the surgical anatomy through schematic representations, cadaveric dissections, and intraoperative photographs and videos, a comprehensive understanding of the approach may still be difficult, especially for less experienced neurosurgeons. In this work, we reported the application of advanced technology for three-dimensional (3D) modeling and visualization of the main neurova-scular structures during vertical hemispherotomy procedures. In the first part of the study, we built a detailed 3D model of the main structures and landmarks involved during each disconnection phase. In the second part, we discussed the adjunctive value of augmented reality systems for the management of the most challenging etiologies, such as hemimegalencephaly and post-ischemic encephalopathy. We demonstrated the contribution of advanced 3D modeling and visualization to enhance the quality of anatomical representation and interaction between the operator and model according to a surgical perspective, optimizing the quality of presurgical planning, intraoperative orientation, and educational training.
- Open Access
- 1 January 2023
- Published by SAGE Publications in Digital Health
- Vol. 9
Magnetic resonance imaging based neurosurgical planning on hololens 2: A feasibility study in a paediatric hospital
Objective: The goal of this work is to show how to implement a mixed reality application (app) for neurosurgery planning based on neuroimaging data, highlighting the strengths and weaknesses of its design. Methods: Our workflow explains how to handle neuroimaging data, including how to load morphological, functional and diffusion tensor imaging data into a mixed reality environment, thus creating a first guide of this kind. Brain magnetic resonance imaging data from a paediatric patient were acquired using a 3 T Siemens Magnetom Skyra scanner. Initially, this raw data underwent specific software pre-processing and were subsequently transformed to ensure seamless integration with the mixed reality app. After that, we created three-dimensional models of brain structures and the mixed reality environment using Unity™ engine together with Microsoft® HoloLens 2™ device. To get an evaluation of the app we submitted a questionnaire to four neurosurgeons. To collect data concerning the performance of a user session we used Unity Performance Profiler. Results: The use of the interactive features, such as rotating, scaling and moving models and browsing through menus, provided by the app had high scores in the questionnaire, and their use can still be improved as suggested by the performance data collected. The questionnaire's average scores were high, so the overall experiences of using our mixed reality app were positive. Conclusion: We have successfully created a valuable and easy-to-use neuroimaging data mixed reality app, laying the foundation for more future clinical uses, as more models and data derived from various biomedical images can be imported.
- 1 January 2023
- Published by SAGE Publications in Digital Health
- Vol. 9
Open AccessMagnetic resonance imaging based neurosurgical planning on hololens 2: A feasibility study in a paediatric hospital
Objective: The goal of this work is to show how to implement a mixed reality application (app) for neurosurgery planning based on neuroimaging data, highlighting the strengths and weaknesses of its design. Methods: Our workflow explains how to handle neuroimaging data, including how to load morphological, functional and diffusion tensor imaging data into a mixed reality environment, thus creating a first guide of this kind. Brain magnetic resonance imaging data from a paediatric patient were acquired using a 3 T Siemens Magnetom Skyra scanner. Initially, this raw data underwent specific software pre-processing and were subsequently transformed to ensure seamless integration with the mixed reality app. After that, we created three-dimensional models of brain structures and the mixed reality environment using Unity™ engine together with Microsoft® HoloLens 2™ device. To get an evaluation of the app we submitted a questionnaire to four neurosurgeons. To collect data concerning the performance of a user session we used Unity Performance Profiler. Results: The use of the interactive features, such as rotating, scaling and moving models and browsing through menus, provided by the app had high scores in the questionnaire, and their use can still be improved as suggested by the performance data collected. The questionnaire's average scores were high, so the overall experiences of using our mixed reality app were positive. Conclusion: We have successfully created a valuable and easy-to-use neuroimaging data mixed reality app, laying the foundation for more future clinical uses, as more models and data derived from various biomedical images can be imported.
- 31 March 2022
- Published by Elsevier in Clinical Neurology and Neurosurgery
- Vol. 217, 107236
“De novo” brain arteriovenous malformation in a child with congenital porto-systemic shunt and multisystemic angiomas
- 31 March 2022
- Published by Elsevier in Clinical Neurology and Neurosurgery
- Vol. 217, 107236
“De novo” brain arteriovenous malformation in a child with congenital porto-systemic shunt and multisystemic angiomas
- 31 October 2021
- Published by Oxford University Press (OUP) in European Journal of Cardio-Thoracic Surgery
- Vol. 62 (1)
Bleeding in patients with continuous-flow left ventricular assist devices: acquired von Willebrand disease or antithrombotics?
OBJECTIVES: To evaluate the competing pro-haemorrhagic contribution of acquired von Willebrand (vW) disease and antithrombotic therapy in patients implanted with continuous-flow left ventricular assist devices (LVADs). METHODS: We compared the extent of vW factor (vWf) degradation [vWf antigen (vWf:Ag)] and a decrease of functional activity of large vWf multimers [vWf collagen binding (vWf:CB)] in LVAD patients who did and did not suffer from bleeding. Data were measured pre-implant, at short-term (t1: 12 months) follow-up. The occurrence of primary bleeding events, as well as bleeding recurrence, was correlated with patient-specific vWf profile and antithrombotic regimen. Indeed, patients were discharged on warfarin (international normalized ratio: 2–2.5) and aspirin, with the latter withhold after a first bleeding episode. RESULTS: Fifty-three patients were enrolled. The median follow-up was 324 (226–468) days. We recorded 25 primary bleeding events (47% of patients). All primary events occurred in patients on warfarin and aspirin. Both vWf:Ag and vWf:CB decreased significantly post-implant (P = 0.0003 and P < 0.0001), and patients showing pathological vWf:CB/vWf:Ag ratio (<0.7) increased progressively over the time of support (pre-implant = 26%, t1 = 58%, t2 = 74%; P < 0.0001). Of note, activity of large vWf multimers of bleeders was significantly lower at t2 with respect to non-bleeders (vWf:CB: 61 (36–115) vs 100 (68–121), P = 0.04; vWf:CB/vWf:Ag ratio: 0.36 (0.26–0.61) vs 0.58 (0.33–0.96), P = 0.04). Despite these marked differences in the vWf profile, following aspirin discontinuation only 3 patients had bleeding recurrence. CONCLUSIONS: Aspirin contributes significantly to haemorrhagic events in the background of acquired vW disease; its discontinuation significantly reduces bleeding recurrence. Clinical trial registration: https://clinicaltrials.gov/ct2/show/NCT03255928; ClinicalTrials.gov Identifier: NCT03255928.
- 31 October 2021
- Published by Oxford University Press (OUP) in European Journal of Cardio-Thoracic Surgery
- Vol. 62 (1)
Bleeding in patients with continuous-flow left ventricular assist devices: acquired von Willebrand disease or antithrombotics?
OBJECTIVES: To evaluate the competing pro-haemorrhagic contribution of acquired von Willebrand (vW) disease and antithrombotic therapy in patients implanted with continuous-flow left ventricular assist devices (LVADs). METHODS: We compared the extent of vW factor (vWf) degradation [vWf antigen (vWf:Ag)] and a decrease of functional activity of large vWf multimers [vWf collagen binding (vWf:CB)] in LVAD patients who did and did not suffer from bleeding. Data were measured pre-implant, at short-term (t1: 12 months) follow-up. The occurrence of primary bleeding events, as well as bleeding recurrence, was correlated with patient-specific vWf profile and antithrombotic regimen. Indeed, patients were discharged on warfarin (international normalized ratio: 2–2.5) and aspirin, with the latter withhold after a first bleeding episode. RESULTS: Fifty-three patients were enrolled. The median follow-up was 324 (226–468) days. We recorded 25 primary bleeding events (47% of patients). All primary events occurred in patients on warfarin and aspirin. Both vWf:Ag and vWf:CB decreased significantly post-implant (P = 0.0003 and P < 0.0001), and patients showing pathological vWf:CB/vWf:Ag ratio (<0.7) increased progressively over the time of support (pre-implant = 26%, t1 = 58%, t2 = 74%; P < 0.0001). Of note, activity of large vWf multimers of bleeders was significantly lower at t2 with respect to non-bleeders (vWf:CB: 61 (36–115) vs 100 (68–121), P = 0.04; vWf:CB/vWf:Ag ratio: 0.36 (0.26–0.61) vs 0.58 (0.33–0.96), P = 0.04). Despite these marked differences in the vWf profile, following aspirin discontinuation only 3 patients had bleeding recurrence. CONCLUSIONS: Aspirin contributes significantly to haemorrhagic events in the background of acquired vW disease; its discontinuation significantly reduces bleeding recurrence. Clinical trial registration: https://clinicaltrials.gov/ct2/show/NCT03255928; ClinicalTrials.gov Identifier: NCT03255928.