Downloaded 20 times
Uploaded byJoel Saltz
Machine Learning and Deep Contemplation of Data
The document discusses various applications of machine learning and data analysis techniques in biomedical fields, focusing on spatio-temporal sensor integration and characterization of complex biological structures. It highlights the importance of generating features, data cleaning, and the use of machine learning for predictive analytics in precision medicine and cancer treatment. Additionally, the document outlines ongoing collaborations and the significance of robust algorithms in analyzing imaging data for improved diagnostics and classification of diseases.
More Related Content
Similar to Machine Learning and Deep Contemplation of Data
Machine Learning and Deep Contemplation of Data
- 1. Machine Learning andDeep Contemplation of Data Joel Saltz Department of Biomedical Informatics Stony Brook University CCDSC October 5, 2016
- 2. From BDEC: “Domain”:Spatio-temporal Sensor Integration, Analysis, Classification • Multi-scale material/tissue structural, molecular, functional characterization. Design of materials with specific structural, energy storage properties, brain, regenerative medicine, cancer • Integrative multi-scale analyses of the earth, oceans, atmosphere, cities, vegetation etc – cameras and sensors on satellites, aircraft, drones, land vehicles, stationary cameras • Digital astronomy • Hydrocarbon exploration, exploitation, pollution remediation • Solid printing integrative data analyses • Data generated by numerical simulation codes – PDEs, particle methods
- 3. Things that Needto be Done with Spatio Temporal Data • Generation of Features • Sanity Checking and Data Cleaning • Qualitative Exploration • Descriptive Statistics • Classification • Identification of Interesting Phenomena • Prediction • Control • Save Data for Later (Compression)
- 4. Precision Medicine MetaApplication • Predict treatment outcome, select, monitor treatments • Reduce inter-observer variability in diagnosis • Computer assisted exploration of new classification schemes • Multi-scale cancer simulations
- 5. Imaging and PrecisionMedicine - Pathomics, Radiomics Identify and segment trillions of objects – nuclei, glands, ducts, nodules, tumor niches … from Pathology, Radiology imaging datasets Extract features from objects and spatio-temporal regions Support queries against ensembles of features extracted from multiple datasets Statistical analyses and machine learning to link Radiology/Pathology features to “omics” and outcome biological phenomena Principle based analyses to bridge spatio-temporal scales – linked Pathology, Radiology studies
- 6. Things that Needto be Done with Spatio Temporal Data • Generation of Features • Sanity Checking and Data Cleaning • Qualitative Exploration • Descriptive Statistics • Classification • Identification of Interesting Phenomena • Prediction • Control • Save Data for Later (Compression)
- 7. Current Driving Applications •Checkpoint Inhibitors – when to use, when to stop • Pathology, Imaging data obtained prior to and during treatment • Integration of “omics”, tissue and imaging to manage treatment • Non Small Cell Lung Cancer, Melanoma, Brain • Virtual Tissue Respository • SEER Cancer Epidemiology • 500K Cancer Patients per year • DOE/NCI pilot involving text • Our co-located companion Virtual Tissue Repository pilot targets SEER images
- 8. Radiomics Decoding tumour phenotype bynoninvasive imaging using a quantitative radiomics approach Hugo J. W. L. Aerts et. Al. Nature Communications 5, Article number: 4006 doi:10.1038/ncomms5006 Features Patients
- 9. Integrative Morphology/”omics” Quantitative Feature Analysis inPathology: Emory In Silico Center for Brain Tumor Research (PI = Dan Brat, PD= Joel Saltz) NLM/NCI: Integrative Analysis/Digital Pathology R01LM011119, R01LM009239 (Dual PIs Joel Saltz, David Foran) J Am Med Inform Assoc. 2012 Integrated morphologic analysis for the identification and characterization of disease subtypes. Pathomics Lee Cooper, Jun Kong
- 10. Things that Needto be Done with Spatio Temporal Data • Generation of Features • Sanity Checking and Data Cleaning • Qualitative Exploration • Descriptive Statistics • Classification • Identification of Interesting Phenomena • Prediction • Control • Save Data for Later (Compression)
- 12. Robust Nuclear Segmentation •Robust ensemble algorithm to segment nuclei across tissue types • Optimized algorithm tuning methods • Parameter exploration to optimize quality • Systematic Quality Control pipeline encompassing tissue image quality, human generated ground truth, convolutional neural network critique • Yi Gao, Allen Tannenbaum, Dimitris Samaras, Le Hou, Tahsin Kurc
- 13.
- 14. Things that Needto be Done with Spatio Temporal Data • Generation of Features • Sanity Checking and Data Cleaning • Qualitative Exploration • Descriptive Statistics • Classification • Identification of Interesting Phenomena • Prediction • Control • Save Data for Later (Compression)
- 15. 3D Slicer Pathology– Generate High Quality Ground Truth
- 16.
- 17. Adjust algorithm parameters,manual fine tuning
- 18. Sanity Check Features RelationshipBetween Image and FeaturesLeveraging Visualization to Aid in Feature Management Step 1: Choose a case from the TCGA atlas (case #20) Step 2: Select two features of interest; X axis (area), Y axis (perimeter) Step 3: Zoom in on region of interest Step 4: Pick a specific nucleus of interest. Each dot represents a single nucleus Step 5: Evaluate the features selected in the context of the specific nucleus and where this nucleus is located within the whole slide image The tool provides visual context for feature evaluation. This technique maps both intuitive features (i.e. size, shape, color) and non-intuitive features (i.e. wavelets, texture) to the ground truth of source images through an interactive web-based user interface. Selected nucleus geolocated within whole slide image Detects elongated nucleus Going from the whole slide data set to selected features and back to the image Adding a visual perspective by using a live web-based interactive tool (http://sbu-bmi.github.io/featurescape/u24/Preview.html)
- 20. Select Feature Pair– dots correspond to nuclei
- 21. Subregion selected –form of gating analogous to flow cytometry
- 22. Sample Nuclei fromGated Region
- 23. Gated Nuclei inContext
- 24.
- 25. Heatmap – DepictsAgreement Between Algorithms
- 26. Things that Needto be Done with Spatio Temporal Data • Generation of Features • Sanity Checking and Data Cleaning • Qualitative Exploration • Descriptive Statistics • Classification • Identification of Interesting Phenomena • Prediction • Control • Save Data for Later (Compression)
- 27. Auto-tuning and featureextraction • Goal – correctly segment trillions of objects (nuclei) • Adjust algorithm parameters • Autotuning – finds parameters that best match ground truth in an image patch • Region template runtime support to optimize generation and management of multi-parameter algorithm results • Eliminates redundant computation, manages locality • Active Harmony – Jeff Hollingsworth!! • Collaboration – George Teodoro, Tahsin Kurc
- 29.
- 30. Performance Optimization 256 nodesof Stampede. Each node of the cluster has a dual socket Intel Xeon E5-2680 processors, an Intel Xeon Phi SE10P co-processor and 32GB RAM.The nodes are inter-connected via Mellanox FDR Infiniband switches.
- 31. Good Bad Test asGood 2916 33 Test as Bad 28 2094 Machine Learning and Quality Critiquing SVM Approach
- 32. Things that Needto be Done with Spatio Temporal Data • Generation of Features • Sanity Checking and Data Cleaning • Qualitative Exploration • Descriptive Statistics • Classification • Identification of Interesting Phenomena • Prediction • Control • Save Data for Later (Compression)
- 33. Feature Explorer -Integrated Pathomics Features, Outcomes and “omics” – TCGA NSCLC Adeno Carcinoma Patients
- 34. Feature Explorer -Integrated Pathomics Features, Outcomes and “omics” – TCGA NSCLC Adeno Carcinoma Patients
- 35. Collaboration with MGH– Feature Explorer – Radiology Brain MR/Pathology Features
- 36. Collaboration with SBURadiology – TCGA NSCLC Adeno Carcinoma Integrative Radiology, Pathology, “omics”, outcome Mary Saltz, Mark Schweitzer SBU Radiology
- 37. Things that Needto be Done with Spatio Temporal Data • Generation of Features • Sanity Checking and Data Cleaning • Qualitative Exploration • Descriptive Statistics • Classification • Identification of Interesting Phenomena • Prediction • Control • Save Data for Later (Compression)
- 38. Classification • Automated orsemi-automated identification of tissue or cell type • Variety of machine learning and deep learning methods • Classification of Neuroblastoma • Classification of Gliomas • Quantification of lymphocyte infiltration
- 39. Classification and Characterizationof Heterogeneity Gurcan, Shamada, Kong, Saltz Hiro Shimada, Metin Gurcan, Jun Kong, Lee Cooper Joel Saltz BISTI/NIBIB Center for Grid Enabled Image Analysis - P20 EB000591, PI Saltz lassification and Characterization of Heterogeneity
- 40. Neuroblastoma Classification FH: favorablehistology UH: unfavorable histology CANCER 2003; 98:2274-81 <5 yr Schwannian Development ≥50% Grossly visible Nodule(s) absent present Microscopic Neuroblastic foci absent present Ganglioneuroma (Schwannian stroma-dominant) Maturing subtype Mature subtype Ganglioneuroblastoma, Intermixed (Schwannian stroma-rich) FH FH Ganglioneuroblastoma, Nodular (composite, Schwannian stroma-rich/ stroma-dominant and stroma-poor) UH/FH* Variant forms* None to <50% Neuroblastoma (Schwannian stroma-poor) Poorly differentiated subtype Undifferentiated subtype Differentiating subtype Any age UH ≥200/5,000 cells Mitotic & karyorrhectic cells 100-200/5,000 cells <100/5,000 cells Any age ≥1.5 yr <1.5 yr UH UH FH ≥200/5,000 cells 100-200/5,000 cells <100/5,000 cells Any age UH ≥1.5 yr <1.5 yr ≥5 yr UH FH UH FH
- 41. Multi-Scale Machine LearningBased Shimada Classification System • Background Identification • Image Decomposition (Multi-resolution levels) • Image Segmentation (EMLDA) • Feature Construction (2nd order statistics, Tonal Features) • Feature Extraction (LDA) + Classification (Bayesian) • Multi-resolution Layer Controller (Confidence Region) No Yes Image Tile Initialization I = L Background? Label Create Image I(L) Segmentation Feature Construction Feature Extraction Classification Segmentation Feature Construction Feature Extraction Classifier Training Down-sampling Training Tiles Within Confidence Region ? I = I -1 I > 1? Yes Yes No No TRAINING TESTING
- 43. Brain Tumor Classification– CVPR 2016
- 44. Combining Information fromPatches
- 45. Brain Tumor ClassificationResults Le Hou, Dimitris Samaras, Tahsin Kurc, Yi Gao, Liz Vanner, James Davis, Joel Saltz
- 46. Tumor Infiltrating Lymphocytequantification • Convolutional neural network to classify lymphocyte infiltration in tissue patches • Convolutional neural network and random forest to classify individual segmented nuclei • Extensive collection of ground truth • Joint work with Emory and TCGA PanCanAtlas Immune group Unsupervised Autoencoder – 100 feature dimensions
- 47. Lymphocyte identification Lymphocytes InfiltrationNo Lymphocyte Infiltration
- 48. Receiver Operating Characteristic– Area Under Curve – 95%
- 49. Lymphocyte Classification HeatMap Trained with 22.2K image patches Pathologist corrects and edits
- 50. Commonalities • Provided quickbut pretty deep dive into aspects of spatio temporal data analytics • Requirements, methods and I think core infrastructure can be shared between disparate application classes • These application classes are definitely data but spatio-temporal aspects are HPC community context friendly • Most of this holds for analysis of scientific program generated data – ORNL Klasky collaborations
- 51. ITCR Team Stony BrookUniversity Joel Saltz Tahsin Kurc Yi Gao Allen Tannenbaum Erich Bremer Jonas Almeida Alina Jasniewski Fusheng Wang Tammy DiPrima Andrew White Le Hou Furqan Baig Mary Saltz Emory University Ashish Sharma Adam Marcus Oak Ridge National Laboratory Scott Klasky Dave Pugmire Jeremy Logan Yale University Michael Krauthammer Harvard University Rick Cummings
- 52. Funding – Thanks! •This work was supported in part by U24CA180924- 01, NCIP/Leidos 14X138 and HHSN261200800001E from the NCI; R01LM011119-01 and R01LM009239 from the NLM • This research used resources provided by the National Science Foundation XSEDE Science Gateways program under grant TG-ASC130023 and the Keeneland Computing Facility at the Georgia Institute of Technology, which is supported by the NSF under Contract OCI-0910735.
- 53.
Editor's Notes
- #41 This is Dr. Shimada’s prognosis classfication