If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Secondary Site Cefalù - Institute of Bioimaging and Molecular Physiology (IBFM) - National Research Council (CNR), Laboratori Nazionali del Sud - National Institute for Nuclear Physics (LNS-INFN)
Dipartimento di Scienze del Suolo, della Pianta e degli Alimenti - Università degli Studi di Bari Aldo Moro, Via G. Amendola, 165/a, 70126 Bari, ItalyIstituto Nazionale di Fisica Nucleare - Sezione di Bari, Via Orabona 4, 70125 Bari, Italy
Istituto Nazionale di Fisica Nucleare - Sezione di Bari, Via Orabona 4, 70125 Bari, ItalyDipartimento Interateneo di Fisica, Università degli Studi di Bari Aldo Moro, Via Orabona 4, 70125 Bari, Italy
A systematic search for papers on applications of AI to medical imaging in Italy was performed.
•
168 research papers were selected 65% using machine learning, 35% deep learning.
•
A rapid increase of interest in AI was observed in the last years.
•
Further collaborations, initiatives and guidelines are needed to develop the research on AI on Imaging.
Abstract
Purpose
To perform a systematic review on the research on the application of artificial intelligence (AI) to imaging published in Italy and identify its fields of application, methods and results.
Materials and Methods
A Pubmed search was conducted using terms Artificial Intelligence, Machine Learning, Deep learning, imaging, and Italy as affiliation, excluding reviews and papers outside time interval 2015–2020. In a second phase, participants of the working group AI4MP on Artificial Intelligence of the Italian Association of Physics in Medicine (AIFM) searched for papers on AI in imaging.
Results
The Pubmed search produced 794 results. 168 studies were selected, of which 122 were from Pubmed search and 46 from the working group. The most used imaging modality was MRI (44%) followed by CT(12%) ad radiography/mammography (11%). The most common clinical indication were neurological diseases (29%) and diagnosis of cancer (25%). Classification was the most common task for AI (57%) followed by segmentation (16%). 65% of studies used machine learning and 35% used deep learning. We observed a rapid increase of research in Italy on artificial intelligence in the last 5 years, peaking at 155% from 2018 to 2019.
Conclusions
We are witnessing an unprecedented interest in AI applied to imaging in Italy, in a diversity of fields and imaging techniques. Further initiatives are needed to build common frameworks and databases, collaborations among different types of institutions, and guidelines for research on AI.
Accurate and early diagnosis and prognosis are essential in many fields of healthcare. Artificial Intelligence (AI) applied to medical images allows for automated disease detection, characterization of histology, stage, or subtype, and patient classification according to therapy outcome or prognosis. It also permits outlining particular regions in the images, quantifying organ volumes, and extracting features from the images which, combined with machine learning algorithms, leads to quantification of image properties or image classification.
In recent years an unprecedented amount of digital imaging data has become available in medicine thanks to digitalization, affordable data storage, and improved imaging techniques. This leads to an unprecedented interest in applications of AI to images which has boosted research efforts of medical physicists (MPs) in Italy.
These efforts however, are more efficient when there is communication, collaboration, sharing of knowledge and common intents in the MP community. For these purposes, the Italian Association of Medical Physics (AIFM) [
Associazione Italiana di Fisica Medica e Sanitaria. Big Data e Intelligenza Artificiale: il gruppo di lavoro AIFM. AIFM 2020. https://www.fisicamedica.it/i-temi-della-fisica-medica/big-data-e-intelligenza-artificiale/big-data-e-intelligenza-artificiale-il-gruppo-di-lavoro-aifm/ (accessed March 1, 2021).
The purpose of this review, performed by the imaging subgroup of AI4MP, is to describe the status of the research in Italy on AI applied to imaging, by systematically analyzing research published in this field in Italy in the last 5 years. This work, besides providing an overview of the fields of application, methods for AI used and the results achieved, will serve to define future goals for the community of MP and facilitate research on AI applied to imaging by MPs in Italy.
Machine learning and deep learning in imaging
Machine learning (ML) is a field of AI algorithms (Fig. 1) which can recognize patterns in medical images by analyzing voxel intensity values or quantitative imaging features, called also “radiomic features”, by identifying their best combination and building a model for classification or regression [
Lella E, Lombardi A, Amoroso N, Diacono D, Maggipinto T, Monaco A, et al. Machine learning and DWI brain communicability networks for Alzheimer’s disease detection. Appl Sci 2020;10. https://doi.org/10.3390/app10030934.
Alongi P, Laudicella R, Stefano A, Caobelli F, Comelli A, Vento A, et al. Choline PET/CT features to predict survival outcome in high risk prostate cancer restaging: a preliminary machine-learning radiomics study. Q J Nucl Med Mol Imaging Off Publ Ital Assoc Nucl Med AIMN Int Assoc Radiopharmacol IAR Sect Soc Of 2020. https://doi.org/10.23736/S1824-4785.20.03227-6.
D’Amico NC, Sicilia R, Cordelli E, Tronchin L, Greco C, Fiore M, et al. Radiomics-based prediction of overall survival in lung cancer using different volumes-of-interest. Appl Sci 2020;10. https://doi.org/10.3390/app10186425.
Electron density and biologically effective dose (BED) radiomics-based machine learning models to predict late radiation-induced subcutaneous fibrosis.
Electron density and biologically effective dose (BED) radiomics-based machine learning models to predict late radiation-induced subcutaneous fibrosis.
A large number of supervised ML algorithms is available, as shown in Table 1. Parametric algorithms make an assumption about the functional form of the function that map covariates to the outcome, then learn a finite number of coefficients for the function from the training data. These algorithms, by using a pre-selected function, are generally faster and easier to interpret. Generalized Linear Model with LASSO [
Botta F, Raimondi S, Rinaldi L, Bellerba F, Corso F, Bagnardi V, et al. Association of a CT-based clinical and radiomics score of non-small cell lung cancer (NSCLC) with lymph node status and overall survival. Cancers 2020;12. https://doi.org/10.3390/cancers12061432.
Fractal dimension of cerebral white matter: A consistent feature for prediction of the cognitive performance in patients with small vessel disease and mild cognitive impairment.
], Ridge or Elastic Net penalty and Logistic Regression (LR) belong to parametric ML. Non-parametric algorithms, by using a number of parameters which is not limited, are usually slower and require larger dataset. These include classification and Regression Trees (CART) [
Lombardi A, Amoroso N, Diacono D, Monaco A, Tangaro S, Bellotti R. Extensive evaluation of morphological statistical harmonization for brain age prediction. Brain Sci 2020;10:10.3390/brainsci10060364.
], and Support Vector Machines (SVM). SVM, based on finding a hyperplane that best divides the data into two classes in the feature space, is among most popular ML algorithm and is employed for both classification [
Electron density and biologically effective dose (BED) radiomics-based machine learning models to predict late radiation-induced subcutaneous fibrosis.
Retico A, Giuliano A, Tancredi R, Cosenza A, Apicella F, Narzisi A, et al. The effect of gender on the neuroanatomy of children with autism spectrum disorders: a support vector machine case-control study. Mol Autism 2016;7:5-015-0067-3. eCollection 2016. https://doi.org/10.1186/s13229-015-0067-3.
Lombardi A, Amoroso N, Diacono D, Monaco A, Tangaro S, Bellotti R. Extensive evaluation of morphological statistical harmonization for brain age prediction. Brain Sci 2020;10:10.3390/brainsci10060364.
Table 1Summary of research papers published in Italy in years 2015–2020 using MLapplied to imaging. For each paper, the first author, year of publication, image modality, software and algorithms used for feature selection and model building, methods for model evaluation and validation, and purpose of the study are reported.
Lella E, Lombardi A, Amoroso N, Diacono D, Maggipinto T, Monaco A, et al. Machine learning and DWI brain communicability networks for Alzheimer’s disease detection. Appl Sci 2020;10. https://doi.org/10.3390/app10030934.
Lombardi A, Amoroso N, Diacono D, Monaco A, Tangaro S, Bellotti R. Extensive evaluation of morphological statistical harmonization for brain age prediction. Brain Sci 2020;10:10.3390/brainsci10060364.
Fractal dimension of cerebral white matter: A consistent feature for prediction of the cognitive performance in patients with small vessel disease and mild cognitive impairment.
Retico A, Giuliano A, Tancredi R, Cosenza A, Apicella F, Narzisi A, et al. The effect of gender on the neuroanatomy of children with autism spectrum disorders: a support vector machine case-control study. Mol Autism 2016;7:5-015-0067-3. eCollection 2016. https://doi.org/10.1186/s13229-015-0067-3.
Battineni G, Chintalapudi N, Amenta F, Traini E. A Comprehensive Machine-Learning Model Applied to Magnetic Resonance Imaging (MRI) to Predict Alzheimer’s Disease (AD) in Older Subjects. J Clin Med 2020;9:10.3390/jcm9072146.
One-class support vector machines identify the language and default mode regions as common patterns of structural alterations in young children with autism spectrum disorders.
Artificial intelligence and radiomics enhance the positive predictive value of digital chest tomosynthesis for lung cancer detection within SOS clinical trial.
A fully automatic, threshold-based segmentation method for the estimation of the Metabolic Tumor Volume from PET images: validation on 3D printed anthropomorphic oncological lesions.
Fanizzi A, Basile TMA, Losurdo L, Bellotti R, Bottigli U, Dentamaro R, et al. A machine learning approach on multiscale texture analysis for breast microcalcification diagnosis. BMC Bioinformatics 2020;21:91-020-3358–4. https://doi.org/10.1186/s12859-020-3358-4.
Giannini V, Rosati S, Regge D, Balestra G. Texture Features and Artificial Neural Networks: A Way to Improve the Specificity of a CAD System for Multiparametric MR Prostate Cancer. In: Kyriacou E, Christofides S, Pattichis CS, editors., Cham: Springer International Publishing; 2016, p. 296–301.
MRI-based radiomics signature for localized prostate cancer: a new clinical tool for cancer aggressiveness prediction? Sub-study of prospective phase II trial on ultra-hypofractionated radiotherapy (AIRC IG-13218).
Recenti M, Ricciardi C, Gìslason M, Edmunds K, Carraro U, Gargiulo P. Machine Learning Algorithms Predict Body Mass Index Using Nonlinear Trimodal Regression Analysis from Computed Tomography Scans. In: Henriques J, Neves N, de Carvalho P, editors. XV Mediterr. Conf. Med. Biol. Eng. Comput. – MEDICON 2019, Cham: Springer International Publishing; 2020, p. 839–46.
The role of carboxy-terminal cross-linking telopeptide of type I collagen, dual x-ray absorptiometry bone strain and Romberg test in a new osteoporotic fracture risk evaluation: A proposal from an observational study.
Ricciardi C, Edmunds KJ, Recenti M, Sigurdsson S, Gudnason V, Carraro U, et al. Assessing cardiovascular risks from a mid-thigh CT image: a tree-based machine learning approach using radiodensitometric distributions. Sci Rep 2020;10:2863-020-59873–9. https://doi.org/10.1038/s41598-020-59873-9.
Usefulness of regional right ventricular and right atrial strain for prediction of early and late right ventricular failure following a left ventricular assist device implant: a machine learning approach.
A machine learning-based approach to directly compare the diagnostic accuracy of myocardial perfusion imaging by conventional and cadmium-zinc telluride SPECT.
Alongi P, Laudicella R, Stefano A, Caobelli F, Comelli A, Vento A, et al. Choline PET/CT features to predict survival outcome in high risk prostate cancer restaging: a preliminary machine-learning radiomics study. Q J Nucl Med Mol Imaging Off Publ Ital Assoc Nucl Med AIMN Int Assoc Radiopharmacol IAR Sect Soc Of 2020. https://doi.org/10.23736/S1824-4785.20.03227-6.
Electron density and biologically effective dose (BED) radiomics-based machine learning models to predict late radiation-induced subcutaneous fibrosis.
Botta F, Raimondi S, Rinaldi L, Bellerba F, Corso F, Bagnardi V, et al. Association of a CT-based clinical and radiomics score of non-small cell lung cancer (NSCLC) with lymph node status and overall survival. Cancers 2020;12. https://doi.org/10.3390/cancers12061432.
Schenone D, Lai R, Cea M, Rossi F, Torri L, Bignotti B, et al. Radiomics and artificial intelligence analysis of CT data for the identification of prognostic features in multiple myeloma. Proc.SPIE, vol. 11314, 2020. https://doi.org/10.1117/12.2548983.
Battineni G, Chintalapudi N, Amenta F, Traini E. A Comprehensive Machine-Learning Model Applied to Magnetic Resonance Imaging (MRI) to Predict Alzheimer’s Disease (AD) in Older Subjects. J Clin Med 2020;9:10.3390/jcm9072146.
]. Random forest (RF), a popular concept in ML, are based on a large set of randomly generated decision trees which are trained individually. After training, the prediction is made for all the individual trees and the most frequently selected class is taken as a final result [
Artificial intelligence and radiomics enhance the positive predictive value of digital chest tomosynthesis for lung cancer detection within SOS clinical trial.
Unsupervised ML techniques can determine patterns in the data which can be used for cathegorization, without the need of user-provided labels. Examples are clustering methods such as k-means [
A fully automatic, threshold-based segmentation method for the estimation of the Metabolic Tumor Volume from PET images: validation on 3D printed anthropomorphic oncological lesions.
Schenone D, Lai R, Cea M, Rossi F, Torri L, Bignotti B, et al. Radiomics and artificial intelligence analysis of CT data for the identification of prognostic features in multiple myeloma. Proc.SPIE, vol. 11314, 2020. https://doi.org/10.1117/12.2548983.
Fanizzi A, Basile TMA, Losurdo L, Bellotti R, Bottigli U, Dentamaro R, et al. A machine learning approach on multiscale texture analysis for breast microcalcification diagnosis. BMC Bioinformatics 2020;21:91-020-3358–4. https://doi.org/10.1186/s12859-020-3358-4.
Giannini V, Rosati S, Regge D, Balestra G. Texture Features and Artificial Neural Networks: A Way to Improve the Specificity of a CAD System for Multiparametric MR Prostate Cancer. In: Kyriacou E, Christofides S, Pattichis CS, editors., Cham: Springer International Publishing; 2016, p. 296–301.
Galbusera F, Niemeyer F, Bassani T, Sconfienza LM, Wilke H-J. Estimating the three-dimensional vertebral orientation from a planar radiograph: Is it feasible? 3rd Int Workshop Spine Load Deform 2020;102:109328. https://doi.org/10.1016/j.jbiomech.2019.109328.
Brunese L, Mercaldo F AUID- ORCID: 0000-0002-9425-1657, Reginelli A, Santone A. Radiomics for Gleason Score Detection through Deep Learning. LID - E5411 [pii] LID - 10.3390/s20185411 [doi]. Sens Basel Switz JID - 101204366 n.d.
Mendizabal A, Tagliabue E, Brunet J-N, Dall’Alba D, Fiorini P, Cotin S. Physics-Based Deep Neural Network for Real-Time Lesion Tracking in Ultrasound-Guided Breast Biopsy. In: Miller K, Wittek A, Joldes G, Nash MP, Nielsen PMF, editors. Comput. Biomech. Med., Cham: Springer International Publishing; 2020, p. 33–45.
]is a class of powerful MLs based on multiple deep layers of neural networks, characterized by hundreds of layers, each of which learns to detect different features of increasing complexity from an image. In contrast to ML, DL doesn’t need to define a priori a set of hand-crafted features, instead constructing its own internal features which are able to describe rich and comprehensive information, thus performing data representation and prediction jointly. CNNs can be used for regression [
Marco B, Leonardo B, Enrico G, Francesco S, Gastone C, Claudia T, et al. Circumventing the Curse of Dimensionality in Magnetic Resonance Fingerprinting through a Deep Learning Approach. ArXiv E-Prints 2018:ar:1811.11477.
Castiglioni I, Ippolito D, Interlenghi M, Monti CB, Salvatore C, Schiaffino S, et al. Artificial intelligence applied on chest X-ray can aid in the diagnosis of COVID-19 infection: a first experience from Lombardy, Italy. MedRxiv 2020:2020.04.08.20040907. https://doi.org/10.1101/2020.04.08.20040907.
Brunetti A, Cascarano GD, De Feudis I, Moschetta M, Gesualdo L, Bevilacqua V. Detection and Segmentation of Kidneys from Magnetic Resonance Images in Patients with Autosomal Dominant Polycystic Kidney Disease. In: Huang D-S, Jo K-H, Huang Z-K, editors., Cham: Springer International Publishing; 2019, p. 639–50.
A comparison between two semantic deep learning frameworks for the autosomal dominant polycystic kidney disease segmentation based on magnetic resonance images.
] tasks. Alternatively, DL networks can be used to extract machine learnt from their layers, which can then be passed to ML algorithms for classification [
Antonio F, Danilo C, Simone S, Michele S, Aurelio U. A multimodal deep network for the reconstruction of T2W MR images. ArXiv E-Prints 2019:ar:1908.03009.
], Recurrent Neural Networks (RNN), allow for unlimited flexibility in extracting information (features and filters) outperforming humans in 2D/3D images analysis.
Instead of training a DL from scratch, architectures can be adapted from already existing and trained architecture such as Googlenet [
]. The training an existing network by fine tuning its parameters and weights for a task which is unrelated to the present goal, called transfer learning, enables to develop efficiently accurate models [
] making it possible to train deeper networks. Generative Adversarial Networks (GANs) transport the generative modeling approach in the context of DL. The idea behind this architecture is to learn to generate new data with the same statistics as the training set, using two neural networks contesting each other in a sort of a zero-sum game, where one agent's gain is another agent's loss. In the clinical context it is worth noting the use of GAN for Adversarial Learning, that is to build carefully targeted attacks to fool a model prediction [
A search for peer-reviewed manuscripts written in English was performed using the PubMed engine. The search was aimed at selecting studies performed in Italian institutions using machine learning applied to medical images. The search strategy was: “((((artificial intelligence) OR (machine learning)) OR (deep learning)) AND (imaging)) AND (italy[Affiliation])) NOT (review[Pubblication Type])”, limited to years 2015–2020.
In a second phase, participants to the AI4FM task-group were asked to select the research papers, not covered in the first search, starting from those published from their institute on AI in imaging.
The participants independently reviewed the selected manuscripts. Studies were considered eligible when a machine learning method was applied to images or to features extracted from images, acquired for medical purpose, and the research was performed by investigators affiliated to insitutes in Italy.
For eligible studies, the AI method, type of AI task (classification, regression, segmentation, registration), algorithm, metric used for evaluation of AI performance, and type of validation, were collected. In order to study the type of institution were the research was conducted, the studies were categorized according to the affiliation of the first and last author, as hospital, university, public research institute (mainly National Institute for Nuclear Physics, INFN and National Research Council, CNR) or private company or foundation. The studies were considered as collaborations among different institutions of different types when the first and last authors came from different categories.
Results of systematic review
The search produced 793 papers. 120 studies were removed as not related to the medical field, 223 had not imaging or AI, 52 were editorials or reviews, 244 had not authors from Italy. 32 were removed as they had a significant overlap with other papers with higher impact factor by the same group. 122 were considered eligible, and 46 papers were manually retrieved by the participants in the group, for a total of 168 studies. These are shown in Table 1 (ML studies) and Table 2 (DL studies), were the clinical indication, algorithms and software platform used for each study are described.
Table 2Summary of research papers published in Italy in years 2015–2020 using deep learning applied to imaging. The table shows for each study the first author and year of publication, the image modality, the software for implementing the deep learning, the type of neural network implemented, its training modality, the metric for model evaluation and validation and the purpose of use of the research.
Marco B, Leonardo B, Enrico G, Francesco S, Gastone C, Claudia T, et al. Circumventing the Curse of Dimensionality in Magnetic Resonance Fingerprinting through a Deep Learning Approach. ArXiv E-Prints 2018:ar:1811.11477.
Antonio F, Danilo C, Simone S, Michele S, Aurelio U. A multimodal deep network for the reconstruction of T2W MR images. ArXiv E-Prints 2019:ar:1908.03009.
Brunese L, Mercaldo F AUID- ORCID: 0000-0002-9425-1657, Reginelli A, Santone A. Radiomics for Gleason Score Detection through Deep Learning. LID - E5411 [pii] LID - 10.3390/s20185411 [doi]. Sens Basel Switz JID - 101204366 n.d.
Mendizabal A, Tagliabue E, Brunet J-N, Dall’Alba D, Fiorini P, Cotin S. Physics-Based Deep Neural Network for Real-Time Lesion Tracking in Ultrasound-Guided Breast Biopsy. In: Miller K, Wittek A, Joldes G, Nash MP, Nielsen PMF, editors. Comput. Biomech. Med., Cham: Springer International Publishing; 2020, p. 33–45.
Galbusera F, Niemeyer F, Bassani T, Sconfienza LM, Wilke H-J. Estimating the three-dimensional vertebral orientation from a planar radiograph: Is it feasible? 3rd Int Workshop Spine Load Deform 2020;102:109328. https://doi.org/10.1016/j.jbiomech.2019.109328.
Castiglioni I, Ippolito D, Interlenghi M, Monti CB, Salvatore C, Schiaffino S, et al. Artificial intelligence applied on chest X-ray can aid in the diagnosis of COVID-19 infection: a first experience from Lombardy, Italy. MedRxiv 2020:2020.04.08.20040907. https://doi.org/10.1101/2020.04.08.20040907.
Brunetti A, Cascarano GD, De Feudis I, Moschetta M, Gesualdo L, Bevilacqua V. Detection and Segmentation of Kidneys from Magnetic Resonance Images in Patients with Autosomal Dominant Polycystic Kidney Disease. In: Huang D-S, Jo K-H, Huang Z-K, editors., Cham: Springer International Publishing; 2019, p. 639–50.
A comparison between two semantic deep learning frameworks for the autosomal dominant polycystic kidney disease segmentation based on magnetic resonance images.
We have observed a rapid increase of research in Italy on artificial intelligence in the last 5 years (Fig. 2). The increase in papers peaked at 155% from 2018 to 2019.
Fig. 2histogram of studies on AI applied to imaging published in Italy included in the systematic review versus years of publication.
Fig. 3a-c show the distribution of studies according to image modality, clinical indication, and institution were the research was undertaken.
Fig. 3Pie charts showing the distribution of clinical indication (a), imaging technique (b) and type of institution (c) for studies on AI applied to imaging in Italy.
Results show that most studies (36%) were from universities followed by hospitals (16%). The most used imaging modality was MRI (44%), followed by CT(11%) ad radiography/mammography (10%). and the most common clinical indications were neurological diseases (29%) and diagnosis of cancer (25%).
Classification was the most common task for AI (57%) followed by segmentation (16%). Among ML algorithms, SVM was by far the most used ML (40.2%). Most of the works retrieved use ML techniques (65% of works).
Used software tools
Among the platforms used for ML (Table 1), the most used are Python ML specific libraries, like the Scikit-Learn library [
Recenti M, Ricciardi C, Gìslason M, Edmunds K, Carraro U, Gargiulo P. Machine Learning Algorithms Predict Body Mass Index Using Nonlinear Trimodal Regression Analysis from Computed Tomography Scans. In: Henriques J, Neves N, de Carvalho P, editors. XV Mediterr. Conf. Med. Biol. Eng. Comput. – MEDICON 2019, Cham: Springer International Publishing; 2020, p. 839–46.
Ricciardi C, Edmunds KJ, Recenti M, Sigurdsson S, Gudnason V, Carraro U, et al. Assessing cardiovascular risks from a mid-thigh CT image: a tree-based machine learning approach using radiodensitometric distributions. Sci Rep 2020;10:2863-020-59873–9. https://doi.org/10.1038/s41598-020-59873-9.
], where ML analyses are made intuitive by a graphical interface in which the tasks in the workflow (e.g. data reading, ML training and ML prediction) are visualised as nodes which can be configured and connected to each other. Commercial packages Matlab (Mathworks, Natick, MA), and SAS (SAS Institute, Cary, NC, USA) include ML and DL packages and graphical interfaces [
Lombardi A, Amoroso N, Diacono D, Monaco A, Tangaro S, Bellotti R. Extensive evaluation of morphological statistical harmonization for brain age prediction. Brain Sci 2020;10:10.3390/brainsci10060364.
], neuralnet for ANN, stats and kohonen for k-means/hierarchical clustering and self-organizing maps. The caret package includes many algorithms implemented in different packages by specifying them as methods of a single training function·H2O is a package for AI in R that offers parallelized implementations of ML [
Retico A, Giuliano A, Tancredi R, Cosenza A, Apicella F, Narzisi A, et al. The effect of gender on the neuroanatomy of children with autism spectrum disorders: a support vector machine case-control study. Mol Autism 2016;7:5-015-0067-3. eCollection 2016. https://doi.org/10.1186/s13229-015-0067-3.
] as shown in Table 2. Complex networks are graphs described by pairs of nodes and links that represent the elements of the system to be modelled and the iterations between the same, respectively, and allow to measure particularly informative topological features.
Mostly, DL (Table 2) have been used for End-to-End learning, i.e. networks learn how to do tasks automatically from raw data provided to only interested task, whereas 7 articles developed a DL model as features extractor for classification [
Marco B, Leonardo B, Enrico G, Francesco S, Gastone C, Claudia T, et al. Circumventing the Curse of Dimensionality in Magnetic Resonance Fingerprinting through a Deep Learning Approach. ArXiv E-Prints 2018:ar:1811.11477.
Castiglioni I, Ippolito D, Interlenghi M, Monti CB, Salvatore C, Schiaffino S, et al. Artificial intelligence applied on chest X-ray can aid in the diagnosis of COVID-19 infection: a first experience from Lombardy, Italy. MedRxiv 2020:2020.04.08.20040907. https://doi.org/10.1101/2020.04.08.20040907.
Brunetti A, Cascarano GD, De Feudis I, Moschetta M, Gesualdo L, Bevilacqua V. Detection and Segmentation of Kidneys from Magnetic Resonance Images in Patients with Autosomal Dominant Polycystic Kidney Disease. In: Huang D-S, Jo K-H, Huang Z-K, editors., Cham: Springer International Publishing; 2019, p. 639–50.
A comparison between two semantic deep learning frameworks for the autosomal dominant polycystic kidney disease segmentation based on magnetic resonance images.
Antonio F, Danilo C, Simone S, Michele S, Aurelio U. A multimodal deep network for the reconstruction of T2W MR images. ArXiv E-Prints 2019:ar:1908.03009.
]. Keras can be run as a stand-alone library or from TensorFlow (from tensorflow version greater than 2), an open-source software library for numeric computation [
Brunese L, Mercaldo F AUID- ORCID: 0000-0002-9425-1657, Reginelli A, Santone A. Radiomics for Gleason Score Detection through Deep Learning. LID - E5411 [pii] LID - 10.3390/s20185411 [doi]. Sens Basel Switz JID - 101204366 n.d.
Mendizabal A, Tagliabue E, Brunet J-N, Dall’Alba D, Fiorini P, Cotin S. Physics-Based Deep Neural Network for Real-Time Lesion Tracking in Ultrasound-Guided Breast Biopsy. In: Miller K, Wittek A, Joldes G, Nash MP, Nielsen PMF, editors. Comput. Biomech. Med., Cham: Springer International Publishing; 2020, p. 33–45.
] are other open source DL frameworks which can be run from Python.
Clinical applications
In the following sections we present the most representative studies on AI applied to imaging in Italy, grouped according to the clinical use of AI.
Neurological applications
The most common neuroimaging issues faced with AI tools are early diagnosis, biomarker identification and understanding the mechanism of development of neurodegenerative [
00167-8&id=gr1.jpg){kind=link}
00167-8&id=gr2.jpg){kind=link}
00167-8&id=gr3.jpg){kind=link}