Machine learning for medical imaging - Daniel Remondini DIFA - INFN Sezione di Bologna
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Machine learning for medical imaging Daniel Remondini DIFA – daniel.remondini@unibo.it INFN Sezione di Bologna
State of the art – “Big data” challenge Many available public data (BIG, i.e. at least Tb-sized): Heterogeneous types of information (imaging, omics, clinical) Databases connect and integrate different data types (genic & metabolic networks, clinical trials, in vitro experiments, catalogues of drug effects and targets) Increased computing and storage power (HPC, GPU , Cloud) Rapid availability and management of data
Public databases and repositories Allow in silico meta-analyses and studies Provide preliminary information before new experiments Transcriptome, Epigenomics, Drugs, Clinical trials, protein structure, …
Big Data for biomedical studies TCIA – the cancer imaging archive TC, PET, NMR data TCGA – The Cancer Genetic Atlas Omics (GEP, NGS, SNP, MET)
TCIA-TCGA integrated imaging/omics databases Collection Cancer type Modalities # TCGA-BLCA Bladder Endothelial Carcinoma CT, CR, MR, PT 106 TCGA-BRCA Breast Cancer MR, MG 139 TCGA-CESC Cervical Squamous Cell Carcinoma and Endocervical Adenocarcinoma MR 54 TCGA-COAD Colon Adenocarcinoma CT 25 TCGA-ESCA Esophageal Carcinoma CT 16 TCGA-GBM Glioblastoma Multiforme MR, CT, DX 262 TCGA-HNSC Head and Neck Squamous Cell Carcinoma CT, MR, PT, RTSTRUCT, RTPLAN, RTDOSE 227 TCGA-KICH Kidney Chromophobe CT, MR 15 TCGA-KIRC Kidney Renal Clear Cell Carcinoma CT, MR, CR 267 TCGA-KIRP Kidney Renal Papillary Cell Carcinoma CT, MR, PT 33 TCGA-LGG Low Grade Glioma MR, CT 199 TCGA-LIHC Liver Hepatocellular Carcinoma MR, CT, PT 97 TCGA-LUAD Lung Adenocarcinoma CT, PT, NM 69 TCGA-LUSC Lung Squamous Cell Carcinoma CT, NM, PT 37 TCGA-OV Ovarian Serous Cystadenocarcinoma CT, MR 143 TCGA-PRAD Prostate Cancer CT, PT, MR 14 TCGA-READ Rectum Adenocarcinoma CT, MR 3 TCGA-SARC Sarcomas CT, MR 5 TCGA-STAD Stomach Adenocarcinoma CT 46 TCGA-THCA Thyroid Cancer CT, PT 6 TCGA-UCEC Uterine Corpus Endometrial Carcinoma CT, CR, MR, PT 58 TCIA – public database of biomedical imaging for many tumours For 21 tumours also omics data are available in TCGA (same samples)
Big Data for clinical studies ADNI – Alzheimer Disease Neuroimaging Initiative (imaging, omics, clinical, biospecimens) ABIDE – Autism Brain Imaging Data Exchange (imaging data on control and autistic samples) IXI – Information eXtraction from Images (600 normal healthy subjects MRI)
Omics multivariate analysis: examples TCGA database OPEN Identification of a T cell gene Glioblastoma 11 Cancer types expression clock obtained by Lung > 2000 samples Adenocarcinoma Breast exploiting a MZ twin design Lung Received: 14 January 2016 Daniel Remondini 1,2, Nathan Intrator3, Claudia Sala1, Michela Pierini4,6, Paolo Garagnani2,4, Squamous Accepted: 1 June 2017 www.nature.com/scientificreports Isabella Zironi 1, Claudio Franceschi5, Stefano Salvioli2,4 & Gastone Castellani1,2 Published: xx xx xxxx Kidney Many studies investigated age-related changes in gene expression of different tissues, with scarce Ovarian Colon agreement due to the high number of affecting factors. Similarly, no consensus has been reached 2017 Drug repurposing www.nature.com/scientificreports/ Endometrial on which genes change expression as a function of age and not because of environment. In this Rectum study we analysed gene expression of T lymphocytes from 27 healthy monozygotic twin couples, with ages ranging over whole adult lifespan (22 to 98 years). This unique experimental design Target identification allowed us OPEN Identification of a T cell gene to identify genes involved in normative aging, which expression changes independently from environmental factors. We obtained a transcriptomic signature with 125 genes, from which expression clock obtained by chronological age can be estimated. This signature has been tested in two datasets of same cell type exploiting a MZ twin design hybridized over two different platforms, showing a significantly better performance compared to random Received: signatures. 14 January 2016 Moreover, the same signature was applied on a dataset from a different cell type Daniel Remondini 1,2, Nathan Intrator3, Claudia Sala1, Michela Pierini4,6, Paolo Garagnani2,4, (human Accepted: 1 June 2017muscle). AIsabella lower performance Zironi 1 was5, Stefano , Claudio Franceschi obtained, Salvioli2,4indicating the1,2possibility that the signature is T & Gastone Castellani Published: xx xx xxxx cell-specific. As a whole our Many studies resultsage-related investigated suggest that changes this in gene approach expression of different can tissues,be withuseful scarce to identify age-modulated agreement due to the high number of affecting factors. Similarly, no consensus has been reached genes. on which genes change expression as a function of age and not because of environment. In this study we analysed gene expression of T lymphocytes from 27 healthy monozygotic twin couples, with ages ranging over whole adult lifespan (22 to 98 years). This unique experimental design allowed us to identify genes involved in normative aging, which expression changes independently from environmental factors. We obtained a transcriptomic signature with 125 genes, from which Aging is a complex phenomenon characterised by decreased fitness and increased risk of diseases, disability and chronological age can be estimated. This signature has been tested in two datasets of same cell type death. All these features are sustained by changes in gene expression, as a response of the cells to the environmen- hybridized over two different platforms, showing a significantly better performance compared to random signatures. Moreover, the same signature was applied on a dataset from a different cell type tal stimuli. Whether this response is programmed and stereotyped or totally random has been (and still is) a puz- (human muscle). A lower performance was obtained, indicating the possibility that the signature is T cell-specific. As a whole our results suggest that this approach can be useful to identify age-modulated zling question for gerontologists. This question stems from the old theoretical dichotomy which has dominated genes. the field of aging studies, that can be summarized in two conflicting positions: “aging is programmed” vs “aging is a random process”1–3Aging . The fact that no gerontogenes (that is, genes whose expression actively induces aging of the is a complex phenomenon characterised by decreased fitness and increased risk of diseases, disability and death. All these features are sustained by changes in gene expression, as a response of the cells to the environmen- organism without any other tal stimuli. apparent Whether benefit) this response is programmed have been found and stereotyped so far or totally random hasdoes zling question for gerontologists. This question stems from the old theoretical dichotomy which has dominated been (andnot exclude still is) a puz- that other (possibly epige- netic) types of control exist, the field of agingsostudies, thethat question can be summarizedis stillin twoopen. conflictingApositions: third“aging possibility is programmed” also exists, vs “aging is that, according to the con- a random process” . The fact that no 1–3 4 gerontogenes (that is, genes whose expression actively induces aging of the ceptualization of Blagosklonny organism without any andotherHall apparent,benefit) aging quasi-programmed, haveisbeen found so far does not exclude that andothershould (possibly epige-be interpreted as a continuation BioPlex Ontocancro of developmental programs ceptualization which, of developmental programs in and theHallpost-reproductive understand, at which, in the post-reproductiveleast in period part, period the of life, loose of life,asloose netic) types of control exist, so the question is still open. A third possibility also exists, that, according to the con- ulation. A possible useful model of Blagosklonny to 4 , aging is quasi-programmed, and should be interpreted presence their strict and finelyof their strict and finely tuned mod- a continuation genetic tuned mod- (or epigenetic) control over ulation. A possible useful model to understand, at least in part, the presence of genetic (or epigenetic) control over Protein-Protein Genes annotated age-related gene expression age-related genechanges and they can be therefore environmental considered is that expression changes is thatof a powerful perturbations, with twins of twins 5 model the further possibility . Indeed, monozygotic (MZ) twins share the same genome 5 monozygotic (MZ) twins share the same genome . Indeed, and they can be therefore considered a powerful model to identify genes whose expression is independent from to identify to cross-validate genesin awhose the data obtained member ofexpression the is independent from Interaction in cancer-related pair with those obtained in the other. Therefore, as MZ twins grow old, it should be possible to observe whether environmental perturbations, some of their geneswith the further change expression accordingly, possibility to cross-validate indicating the presence of some kind of geneticthe control data over obtained in a member of the pair with those obtained in the other. Therefore, gene expression as timeMZseries intwins grow old, brain this phenomenon, or rather if changes are totally private (not shared by the two members of the twin couple). Until today, a plethora of studies analyzed different tissues (including it should areas, adipose be possible to observe whether Network pathways some of theirFigure 2.andon twin genes studies change Plot expressionofbeenMZ couple accordingly, age of different age indicating ofvs. , but not in twins.the age estimated On thepresence case-control studies . by ridge regression with other side, geneof some kind of genetic control over 125-gene signature to predict tissue skeletal muscle) from subjects 6–14 expression pairs have performed so far in limited number old subjects 15 , or in 16, 17 this phenomenon, or rather if changes are totally private (not shared by the two members of the twin couple). Until shows today, a plethora the estimation of studies Departmentanalyzed 1 of sics an gene obtained stronom expression ni ersit of o o time for na o o series theIta validation na 40126 in 2 .different Inter epartmenta tissues set (i.e. all enter (including . the brain twins areas, from adipose training). 3 a ani ni ersit of o o na o o na 40126 Ita . Department of omputer cience act ciences acu t tissue and skeletal muscle) frome subjects of different age6–14 , butannot int twins. e icine On ni ersitthe other side, gene expression sample age from blood cells 4 e i ni ersit i Israe . Department of perimenta Dia nostic pecia Nature Communications 2018 studies on twin pairs haveresent a been performed e eneration a so far in limited Institute number o i i a- utti of i ooldi Ort subjects opae ic Institute, or in case-control studies . of o o na o o na 40138 Ita . I 5 Institute of euro o ica ciences of o o na o o na 40124 Ita15. 16, 17 6 ress: one orator esearc ia i ar iano 1/10 40136 o o na Ita . Danie emon ini an at an Intrator contri ute e ua to t is wor . tefano a io i an astone aste ani oint super ise t is wor . orrespon ence an re uests for materia s s ou e a resse to D. . emai : anie .remon ini uni o.it) 2018 1 Department of sics an stronom ni ersit of o o na o o na 40126 Ita . 2Inter epartmenta enter . Scientific RepoRts | 7: 6005 | DOI:10.1038/s41598-017-05694-2 1 a ani ni ersit of o o na o o na 40126 Ita . 3Department of omputer cience act ciences acu t ARTICLE e i nifor ersit the evalidation dataset obtained i Israe . 4Department of perimenta fromDiathe twin nostic an couple pecia t splitting, e icine ni asersitdescr of o o napared o o nawith 40138the Ita results 5 . I obtained Institute of in euro for a signature based on 71 methylation 21 o ica ciences of o o na o o na 40124 Ita . DOI: 10.1038/s41467-018-06992-7 OPEN 6 resent a ress: one e eneration a orator esearc Institute o i i a- utti i o i Ort opae ic Institute ia i ar iano and1/10 R= 0.91oin 40136 o na theItavalidation . Danie emon set. As aatfirst ini an test for an Intrator contrithe utegoodness e ua to t isofworour. sig Network integration of multi-tumour omics data tefano adataset io i an 100,000 astone asterandomly chosen ani oint super ise t signatures of 125 is wor . orrespon ence probes, and an re uests applied for materia s th regression parameters on one split dataset and validating it on the other. W s ou e a resse to D. . emai : anie .remon ini uni o.it) suggests novel targeting strategies Pearson’s R values, with an average of 0.75 and a SD of 0.09. Remarkably, on ARTICLE Scientific RepoRts | 7: 6005 | DOI:10.1038/s41598-017-05694-2 forming the optimal signature (corresponding to 0.02% of the random sig 1 Ítalo Faria do Valle1,2, Giulia Menichetti3, Giorgia Simonetti 4, Samantha Bruno4, Isabella Zironi 1, high regression performance of the optimal age signature. Concerning pos DOI: Danielle Fernandes Durso4,5, José C.M. Mombach OPEN 10.1038/s41467-018-06992-7 6, Giovanni Martinelli4,7, Gastone Castellani1 & sion values of these 125 genes, we did not observe any significant difference Daniel Remondini 1 samples, after correcting for multiple test analysis (data not shown). This sug
Machine Learning Huge amount of data: data-driven approaches Machine Learning: usage of advanced algorithms for data analysis (e.g. image analysis) 1) Unsupervised methods: a) data clustering (e.g. ROI segmentation) b) feature extraction (e.g. texture features) 2) Supervised methods: the algorithm uses known information (e.g. reference samples, standards) for a) sample classification b) parameter regression (e.g. risk score, age)
Artificial Intelligence and Deep Learning Some machine learning techniques take inspiration from anatomical and functional structures of the brain (i.e. visual cortex, known since the ’60s) Layered (modular, organized) Hierarchical (from contours to shapes) Somatotopic (organization) Hubel & Wiesel J Physiol 160, 1962
Artificial Intelligence and Deep Learning The functional units of Neural Networks take inspiration from neurons (since 1957 Rosenblatt’s perceptron)
Deep Neural Network for Machine Learning - supervised Supervised methods (classification, regression): Feedforward Deep Networks FDN FDN is trained with examples, and generalizes to unseen data My thesis (1996): 4 neurons, 386 Intel CPU, 20 Mb HD Actual FDN: 105-106 neurons, multi-core CPU & GPU, Tb HD (& RAM)
Deep Neural Network for Machine Learning - unsupervised Unsupervised feature extraction: Convolutional Neural Networks CNN CNN layers extract a hierarchy of features (from contours to shapes)
Deep Neural Network for Machine Learning Typical DNN architecture: encoding (CNN) + task (FDN)
Machine Learning for NMR data: examples Data processing: 1) NMR fingerprinting 2) QSM processing Image analysis: 1) Automated segmentation 2) Quantification (feature & texture analysis) 3) Image quality enhancement (super-resolution)
Fingerprinting Associate vectors of features for each MRI ”pixel” (training set) with specific values (B, T1, T2, …) Original strategy [Ma et al, Nature 495, 2013]: define a ”dictionary” of feature/value associations Our strategy: train a DL network to discriminate the feature vectors and reliably associate the physical NMR parameters T1 B1 T2 D Off Barbieri, … Remondini arXiv:1811.11477
Quantitative Susceptibility Mapping Transform Phase data into Magnetic susceptibility data c Analytical function has singularities (”magic angle”) Deep Learning can learn the transform from ”good”examples (our case, reconstructions obtained with other approaches) and overcome singularities Cristiana Fiscone’s Tomorrow C15 talk In collaboration with Prof. R. Bowtell Sir Peter Mansfield Institute, Nottingham UK
Feature extraction & analysis Many observables can be extracted from single pixels (or larger patches) of MRI images - Graylevel histogram - Texture features (based on spatial and intensity proximity) - Segmented region parameters (eccentricity, complexity, fractal dim, …) Each sample is mapped into a high-dimensional feature space N = 102-103 Feature vectors can be used for - low-dimensional visualization - Supervised and unsupervised machine learning
Visualization Several tecniques can be used for low-dimensional reduction and visualization (in 2-3 dim) - PCA/SVD ”family” of methods REPORTS tion to geodesic distance. For faraway points, X. Two simple methods are to connect each The final step applies classical MDS to geodesic distance can be approximated by point to all points within some fixed radius !, the matrix of graph distances DG " {dG (i, j)}, - ISOMAP (network-geodetics) adding up a sequence of “short hops” be- tween neighboring points. These approxima- tions are computed efficiently by finding or to all of its K nearest neighbors (15). These neighborhood relations are represented as a weighted graph G over the data points, with constructing an embedding of the data in a d-dimensional Euclidean space Y that best preserves the manifold’s estimated intrinsic - SNE (Stochastic Neighbor Embedding) shortest paths in a graph with edges connect- edges of weight dX(i, j) between neighboring geometry (Fig. 3C). The coordinate vectors yi ing neighboring data points. points (Fig. 3B). for points in Y are chosen to minimize the The complete isometric feature mapping, In its second step, Isomap estimates the cost function or Isomap, algorithm has three steps, which geodesic distances dM (i, j) between all pairs E ! !#$D G % " #$D Y %! L 2 (1) are detailed in Table 1. The first step deter- of points on the manifold M by computing mines which points are neighbors on the their shortest path distances dG (i, j) in the where DY denotes the matrix of Euclidean manifold M, based on the distances dX (i, j) graph G. One simple algorithm (16) for find- distances {dY (i, j) " !yi & yj!} and !A! L2 SNE 2-d representation of hand-written ISOMAP 2-d representation of face images between pairs of points i, j in the input space ing shortest paths is given in Table 1. the L2 matrix norm '(i, j A2i j . The # operator M H digit images VAN DER AATEN AND INTON Fig. 1. (A) A canonical dimensionality reduction problem from visual perception. The input consists of a sequence of 4096-dimensional vectors, rep- resenting the brightness values of 64 pixel by 64 pixel images of a face rendered with different poses and lighting directions. Applied to N " 698 raw images, Isomap (K " 6) learns a three-dimen- 0 sional embedding of the data’s intrinsic geometric 1 structure. A two-dimensional projection is shown, 2 with a sample of the original input images (red 3 circles) superimposed on all the data points (blue) and horizontal sliders (under the images) repre- 4 senting the third dimension. Each coordinate axis 5 of the embedding correlates highly with one de- 6 gree of freedom underlying the original data: left- 7 right pose (x axis, R " 0.99), up-down pose ( y 8 axis, R " 0.90), and lighting direction (slider posi- tion, R " 0.92). The input-space distances dX(i,j ) 9 given to Isomap were Euclidean distances be- tween the 4096-dimensional image vectors. (B) Isomap applied to N " 1000 handwritten “2”s from the MNIST database (40). The two most significant dimensions in the Isomap embedding, shown here, articulate the major features of the “2”: bottom loop (x axis) and top arch ( y axis). Input-space distances dX(i,j ) were measured by tangent distance, a metric designed to capture the invariances relevant in handwriting recognition (41). Here we used !-Isomap (with ! " 4.2) be- cause we did not expect a constant dimensionality to hold over the whole data set; consistent with this, Isomap finds several tendrils projecting from the higher dimensional mass of data and repre- senting successive exaggerations of an extra stroke or ornament in the digit. (a) Visualization by t-SNE.
Super resolution Average test PSNR (dB) A DNN can learn to ”improve” image quality (resolution) from an adequate training set Results of a DNN trained on natural images 4 Number of backprops feature maps feature maps of low-resolution image of high-resolution image Low-resolution High-resolution image (input) image (output) Original Original / PSNR / PSNR Bicubic Bicubic / 24.04 / 24.04 dB dB Patch extraction Non-linear mapping Reconstruction and representation . 2. Given a low-resolution image Y, the first convolutional layer of the SRCNN extracts a set of feature maps. The cond layer maps these feature maps nonlinearly to high-resolution patch representations. The last layer combines predictions within a spatial neighbourhood to produce the final high-resolution image F (Y). kernel size c ⇥ f1 ⇥ f1 . The output is composed of Here W3 corresponds to c filters of a size n2 ⇥ f3 ⇥ f3 , feature maps. B1 is an n1 -dimensional vector, whose and B3 is a c-dimensional vector. SC / SC 25.58 / 25.58 dB dB SRCNN SRCNN / 27.95 / 27.95 dB dB ch element is associated with a filter. We apply the If the representations of the high-resolution patches ctified Linear Unit (ReLU, max(0, x)) [33] on the filter are in the image domain (i.e.,we can simply reshape each ponses4 . representation to form the patch), Fig. 1. The we expect proposed Super-Resolution Convolution that the filters act like an averaging filter; if Neural Network (SRCNN) surpasses the bicubic baselin .2 Non-linear mapping of the high-resolution patches are in some the representations with other just domains Dong et al., arXiv:1501.00092v3 a few training iterations, and outperforms th e first layer extracts an n1 -dimensional feature for (e.g.,coefficients in terms of some bases), we expect that ch patch. In the second operation, we map each of sparse-coding-based W3 behaves like first projecting the coefficients onto the method (SC) [50] with modera se n1 -dimensional vectors into an n2 -dimensional image domain and then averaging. In training. The either way, W3performance is may be further improved w e. This is equivalent to applying n2 filters which have a set of linear filters. more training iterations. More details are provided rivial spatial support 1 ⇥ 1. This interpretation is only
Classification & regression Machine learning techniques (including DL) can be used to classify samples or to regress parameters (e.g. tumour grade, age): - Partial Least Squares - Support Vector Machine - Discriminant Analysis - K-Nearest Neighbour - Ridge regression - LASSO - …
Daniel Remondini Department of Physics and Astronomy – DIFA INFN Sezione di Bologna – AIM initiative (Artificial Intelligence in Medicine) daniel.remondini@unibo.it www.unibo.it
Per quanto concerne i moderatori, relatori, formatori, tutor, docenti è richiesta dall’Accordo Stato-Regioni vigente apposita dichiarazione esplicita dell’interessato, di trasparenza delle fonti di finanziamento e dei rapporti con soggetti portatori di interessi commerciali relativi agli ultimi due anni dalla data dell’evento. La documentazione deve essere disponibile presso il Provider e conservata per almeno 5 anni. Dichiarazione sul Conflitto di Interessi Il sottoscritto DANIEL REMONDINI in qualità di: □ relatore dell’evento “X CONGRESSO AIRMM - RISONANZA MAGNETICA IN MEDICINA 2019: DALLA RICERCA TECNOLOGICA AVANZATA ALLA PRATICA CLINICA” Milano, 28-29 marzo 2019 da tenersi per conto di Biomedia srl Provider n. 148, ai sensi dell’Accordo Stato-Regione in materia di formazione continua nel settore “Salute” (Formazione ECM) vigente, Dichiara X che negli ultimi due anni NON ha avuto rapporti anche di finanziamento con soggetti portatori di interessi commerciali in campo sanitario
You can also read