Brain-on-Cloud for automatic diagnosis of Alzheimer’s disease from 3D structural magnetic resonance whole-brain scans☆
Introduction
Dementia is affecting around fifty million people worldwide and Alzheimer’s Disease (AD) is the most predominant form [1], contributing up to 70% of all dementia cases as reported by the World Health Organization (WHO)2. AD is a cognitive disorder that begins with mild memory losses and worsens progressively until death, as it damages brain cells irreversibly, ending up with the destruction of the brain area that controls cardiac and respiratory functions [1], [2]. Age may significantly affect the evolution of the syndrome: the prevalence of AD after 85 years of age (up to 35%) is estimated to be higher than in other age groups [3], [4]. However, AD is not an exclusive consequence of biological ageing. According to the WHO, the onset of symptoms before 65 years of age accounts for up to 9% of all AD cases. Anatomically, as neurons are injured, connections between neurons may break down and many brain regions begin to shrink dramatically [5], [6]. Even relatively early in its clinical expression, brain atrophy caused by AD targets mainly the cerebral cortex and the anterior hippocampal regions that are involved in thinking, reasoning and keeping new memories [4], [7], [8].
The diagnosis of AD requires careful medical evaluations, including anamnesis, neuropsychological tests, such as Mini Mental State Examination (MMSE), and other neurobiological exams [5]. In addition, neuroimaging data are extensively used as diagnostic support [6]. Structural Magnetic Resonance Imaging (sMRI) is widely exploited for the investigation of progressive neurological impairment [9], [10]. It offers a painless and non-invasive method of analyzing the anatomical changes of the brain, combining radio waves and strong magnetic fields, in order to guarantee high level of spatial resolution [11], [12]. The clinical utility of sMRI in distinguishing AD from other cognitive disorders is well established [13], especially when using T1-weighted (T1w) sMRI scans [14], [15]. T1w images are useful to analyze the brain structure from an anatomical point of view, reliably differentiating between the gray and white matters [16]. Since strong T1 contrast is present between fluid and more solid anatomical structures, T1w images are more suitable for the morphological assessment of the brain anatomy [16]. Thus, the presence of cortical and hippocampal atrophy caused by AD can be appreciated easily from a T1w sMRI scan [14].
Despite the available diagnostic tools, the diagnosis of AD is still very difficult due to similar symptoms with other cognitive disorders. In clinical practice, AD diagnosis can be confirmed only after the patient’s death by means of a postmortem examination of the brain tissue [17]. Furthermore, an effective cure to reverse or stop AD progression has not been identified, yet [17]. Nevertheless, given the disproportionate aging of the population, AD-related socio-economic impact is continuing to rise [18]. Hence, AD diagnosis is crucial, as a timely therapeutic intervention especially during the initial stages of the syndrome appears to have a positive impact on both the progression of symptoms and the quality of life of diseased subjects [1], [2], [19]. In this regard, Artificial Intelligence (AI)-guided computer-aided systems have been widely parsed to automatically diagnose AD, especially using Deep Learning (DL) algorithms [20].
DL has been one of the crucial factors for the success of AI in the medical field [21], [22]. Compared with conventional Machine Learning (ML) algorithms, DL has multiple advantages in analyzing medical images, presenting high power in identifying complex structures and in automatizing feature extraction [7]. Indeed, DL is able to adaptively learn from the data (i.e., fully data-driven process), obtaining the optimal representation of the problem, without relying on handcrafted features [1], [21]. Handcrafted feature extraction is difficult and time-consuming, especially due to the complexity of the diagnostic problem and the difficulty to model prior knowledge completely [17], [23]. Moreover, handcrafted features potentially lead to non-optimal diagnostic results, as they may not be well coordinated with the classifiers used [10]. Therefore, DL algorithms are usually better suitable than classical ML approaches for generalizing even under slight anatomical changes, like the ones caused by AD [13], [24].
To the best of our knowledge, our manuscript has so far been the first and only to propose an end-to-end framework, named ConvLSTM4AD,3 leveraging exclusively on a Convolutional Long Short-Term Memory (ConvLSTM)-based neural network to investigate the presence of AD [1]. Despite the promising results, ConvLSTM4AD was designed to work with only 5 slices per scan. Furthermore, neither the impact of data handling techniques nor the impact of model parameter optimization were analyzed there. Thus, the motivation behind the hereby presented study is to find a more robust yet lightweight 3D framework, named Brain-on-Cloud, for automatically detecting AD from 3D sMRI whole-brain scans on cloud. In this regard, the focus is on improving the end-to-end ConvLSTM-based model dedicated to efficient learning of AD features while overcoming the main limitations raised in [1] by: increasing the cardinality of 3D sMRI scans used to feed the neural network; increasing the number of analyzed slices per scan; automatizing the entire workflow; and conducting in-depth studies, including the impact of different data handling techniques, the impact of hyper-parameter selection on the performance of the model, and the investigation of the diagnostic performance of Brain-on-Cloud in relation to age, sex and MMSE. Additionally, the entire source code of Brain-on-Cloud is available on GitHub under copyright, to ensure its full reproducibility4.
Section snippets
Literature review
Discovering an algorithm able to automatically classify the anatomical brain changes caused by AD is an interesting research topic for the scientific community.
DL algorithms for automatic AD diagnosis mainly focused on semi-supervised learning algorithms to make full use of both labelled and unlabelled sMRI data, and supervised learning algorithms to make use of labelled sMRI data only. As for semi-supervised learning algorithms, Yu et al. [15] were the first to propose a tensor-train,
Brain-on-Cloud
With the aim to improve our end-to-end ConvLSTM-based neural network, four serial experiments were conducted, adding in each experiment new improvements in relation to the previous one. In the first experiment, only intensity normalization and automated cropping were performed. In the second and third experiments, scan registration and brain extraction were respectively added, as they are the distinctive steps in the analysis of sMRI brain scans [26]. In the fourth experiment, data augmentation
Results
Table 1 reports the performance in classifying AD of the four experiments and of the tuned best experiment, across the five splits of the stratified shuffle-split cross validation scheme, on OASIS-3 test data. Performance is given in terms of average ACC, SP, SE, F1-S and AUC, together with the respective standard deviation. Fig. 4 focuses on the ROC curve and AUC values of the four experiments and of the tuned best experiment, across the five splits of the stratified shuffle-split cross
Discussion
This study proposed a robust yet lightweight 3D framework, Brain-on-Cloud, for automatically classifying AD from 3D sMRI whole-brain scans using a scalable GPU cloud service to ensure its public availability under copyright and, as consequence, a complete reproducibility of the entire algorithm. This outcome was obtained by improving our end-to-end ConvLSTM-based model dedicated to efficient learning of AD features by integration with a set of data handling techniques in addition to the tuning
Conclusion
This study demonstrated that Brain-on-Cloud represents a reliable, efficient, lightweight and easily-reproducible method for automatic diagnosis of AD from 3D sMRI whole-brain scans. As Brain-on-Cloud performs well without segmenting the brain into its portions, it can also be applied to other neurological disorders using volumetric whole-brain data, as proven by Tomassini et al. [39]. There, the applicability of Brain-on-Cloud was extended to another neurodegenerative disorder (i.e.,
CRediT authorship contribution statement
Selene Tomassini: Conceptualization, Data curation, Methodology, Software, Writing – original draft. Agnese Sbrollini: Methodology, Writing – review & editing. Giacomo Covella: Software. Paolo Sernani: Writing – review & editing. Nicola Falcionelli: Writing – review & editing. Henning Müller: Project administration, Supervision. Micaela Morettini: Project administration, Supervision. Laura Burattini: Project administration, Supervision. Aldo Franco Dragoni: Project administration, Supervision.
Declaration of Competing Interest
The authors affirm that this manuscript is an honest, accurate and transparent account of the study being reported, that no important aspects of the study have been omitted, and that any discrepancies from the study as planned (and, if relevant, registered) have been explained. The authors declare no relationships with other people or organizations that could inappropriately bias the work. The authors confirm that neither the manuscript nor any parts of its content are currently under
Acknowledgment
The authors would like to thank the head of neurosurgery of Azienda Ospedaliero Universitaria Ospedali Riuniti di Ancona, dr. Roberto Trignani, for his clinical support in the development of the proposed framework.
References (39)
- et al.
Convolutional neural networks for classification of Alzheimer’s disease: Overview and reproducible evaluation
Med. Image Anal.
(2020) - et al.
Mapping progressive brain structural changes in early Alzheimer’s disease and mild cognitive impairment
Neuropsychologia
(2008) - et al.
Automated classification of Alzheimer’s disease and mild cognitive impairment using a single MRI and deep neural networks
NeuroImage
(2019) - et al.
An end-to-end 3D ConvLSTM-based framework for early diagnosis of Alzheimer’s disease from full-resolution whole-brain sMRI scans
IEEE International Symposium on Computer-Based Medical Systems
(2021) - et al.
Dual-functional neural network for bilateral hippocampi segmentation and diagnosis of Alzheimer’s disease
Int. J. Comput. Assist. Radiol. Surg.
(2020) - et al.
Alzheimer’s disease detection through whole-brain 3D-CNN MRI
Front. Bioeng. Biotechnol.
(2020) - et al.
Early diagnosis of Alzheimer’s disease: A neuroimaging study with deep learning architectures
IEEE Conference on Computer Vision and Pattern Recognition
(2018) - et al.
DeepAD: Alzheimer’s disease classification via deep convolutional neural networks using MRI and fMRI
BioRxiv
(2016) - et al.
A comprehensive study of Alzheimer’s disease classification using convolutional neural networks
arXiv preprint arXiv:1904.07950
(2019) - et al.
A data augmentation-based framework to handle class imbalance problem for Alzheimer’s stage detection
IEEE Access
(2019)
Hierarchical fully convolutional network for joint atrophy localization and Alzheimer’s disease diagnosis using structural MRI
IEEE Trans. Pattern Anal. Mach. Intell.
Image-based state-of-the-art techniques for the identification and classification of brain diseases: A review
Med. Biol. Eng. Comput.
Convolutional LSTM: A deep learning approach for dynamic MRI reconstruction
IEEE International Conference on Trends in Electronics and Informatics
Volumetric convolutional neural network for Alzheimer detection
IEEE International Conference on Trends in Electronics and Informatics
Tensorizing GAN with high-order pooling for Alzheimer’s disease assessment
IEEE Trans. Neural Netw. Learn. Syst.
How MRI makes the brain visible
Make Life Visible
Classification of Alzheimer’s disease from MRI data using an ensemble of hybrid deep convolutional neural networks
IEEE International Midwest Symposium on Circuits and Systems
The public health impact of Alzheimer’s disease, 2000–2050: Potential implication of treatment advances
Annu. Rev. Public Health
Alzheimers disease, the importance of early detection
Eur. Neurol. Rev.
Cited by (8)
On-cloud decision-support system for non-small cell lung cancer histology characterization from thorax computed tomography scans
2023, Computerized Medical Imaging and GraphicsAlzheimer's disease progression detection based on optical fluence rate measurements using alternative laser wavelengths
2024, International Journal for Numerical Methods in Biomedical EngineeringArtificial Cognition for Detection of Mental Disability: A Vision Transformer Approach for Alzheimer’s Disease
2023, Healthcare (Switzerland)A Double-Stage 3D U-Net for On-Cloud Brain Extraction and Multi-Structure Segmentation from 7T MR Volumes
2023, Information (Switzerland)Analyzing global features of magnetic resonance images in widespread neurodegenerative diseases: new hope to understand brain mechanism and robust neurodegenerative disease diagnosis
2023, Medical and Biological Engineering and Computing
- ☆
The research presented here was supported by the projects “Using 3D convolutional neural networks for the identification of lung cancer histological types directly from computed tomography scans” funded by the Cariverona Foundation, Italy, and “Deep learning for early medical diagnosis: A novel methodology for different clinical scenarios” funded by the Department of Information Engineering, Engineering Faculty, Universit Politecnica delle Marche, Ancona, Italy.
- 1
Equal contribution.