Deep Learning Image Classification for Histopathology.pptx
Transcription
Deep Learning Image Classification for Histopathology.pptx
Deep Learning Image Classification for Histopathology Boris Murmann, Sean Fischer, Elaina Chai November 19, 2015 2 ImageNet Large Scale Visual Recognition Challenge http://image-net.org http://devblogs.nvidia.com/parallelforall/mocha-jl-deep-learning-julia/ 3 Best-Performing Topology: Convolutional Neural Network Simple Example: Handwritten Digits http://parse.ele.tue.nl/cluster/2/CNNArchitecture.jpg § Training: Use massive amounts of data to find the parameter set that achieves the best possible fit § Testing: Classify input data using the trained model parameters › Our main interest 4 Software versus Hardware Camp www.qualcomm.com/zeroth § Cloud computing versus mobile computing § CPU/GPU clusters versus custom chips § Virtually unconstrained power versus severely constrained power/energy 5 “Brain-Inspired” Computing Hardware Degree of brain inspiration 6 Memory Signaling Computation Analog Analog Spikes Analog Carver Mead, ca. 1990 Digital Digital Spikes Digital IBM True North, 2014 “Sloppy” Digital Mixed-signal Mixed-signal à Our interest Digital Digital Digital DianNao, 2014 Example Application: Lane Tracking Unlabelled Input Image 7 Output Image with Label Points Example Network Structure Huge number of coefficients Power ~ 100W on a GPU board A. Coates 2014 8 Far to high for practical applications Custom Machine Learning Chip with Embedded Memory Area: 67.7 mm2 § Read from 256-bit, 36MB eDRAM: ~19 pJ § Remaining issue: Can’t fit enough memory on chip § Ongoing work on “network compression” will solve this issue Reference: Y. Chen, MICRO 2014 9 Pushing Further: Mixed-Signal Techniques for Machine Learning Low-swing interconnect ConvNet Training Trained parameters Bit flips and quantization error Software 10 Hardware MixedSignal Compute Blocks Low Voltage Memory w/ error Deterministic coupling between bits Charge Domain Dot Product Kernel § Externally digital, internally analog compute block § Amenable for digital CMOS VLSI integration § ADC energy amortized over several multipliers QP § Multiply via multi-phase charge redistribution § Add via passive charge sharing among multipliers SIGN + 8Cu Cu vOD 4Cu SIGN 2Cu Cu − QN 11 2Cu VDD/2 8Cu § Small unit caps < 1fF 4Cu 16x 8-bit MAC Design in 28nm CMOS § Energy stored in capacitors essentially negligible § Total energy dominated by ADC and control wires/switches § Next step is to work on low-swing control and improved ADC 12 The Machine Learning Team The Bio “Team” Architecture Memory Signaling ALU ALU VX∙Y DX,1 DY,1 DX,2 DY,2 ADC DX,N-1 DY,N-1 DX,N DY,N Mixed-signal compute 13 DOUT,X∙Y § § § Grade III Grade II Grade I Histological Features for Cancer Prognosis Clear cellular differentiation, Consistent cell size and shape Normal mitotic activity Subjective curve § § § No clear cellular differentiation Irregular cell size and shape Accelerated mitotic activity D.H. Patey, R.W. Scarff, The position of histology in the prognosis of carcinoma of the breast. Lancet 211, 801-804 (1928) 14 Manual Inspection Remains the Gold Standard Relationship between histological grade and overall survival in 1830 breast cancer patients Grade I Log-Rank Test % Survival Grade II Grade III χ2 = 198.06 P < 0.0001 Time [years] C.W. Elston, I.O. Ellis, Pathological prognostic factors in breast cancer I. The value of histological grade in breast cancer: Experience from a large study with long-term follow-up. Histopathology 19, 403-410 (1991) 15 The Rise of Digital Histopathology Whole Slide Scanner Stanford Tissue Microarray Database (http://tma.stanford.edu) MikroScan D2 Array Block TA-274 Breast #07b VGH-011A v3 (slice 1.00) - 10395 Growing repositories of digitized tissue samples enables machine learning in histopathology. 16 Machine Learning Uncovering New Prognostic Features Training Images (NKI) Testing Images (VGH) Model Prediction C-Path Model on VGH (n=286) Alive Training Deceased C-Path Survival Model New Prognostic Features Alive Log-Rank P = 0.001 Low-Risk High-Risk Years A. H. Beck, A. R. Sangoi, S. Leung, R. J. Marinelli, T. O. Nielsen, M. J. van de Vijver, R. B. West, M. van de Rijn, D. Koller. "Systematic analysis of breast cancer morphology uncovers stromal features associated with survival." Science translational medicine 3.108 (2011) 17 Eight New Epithelial Features for Cancer Prognosis Epithelial tissue lines any surface in organs and blood vessels throughout the body. Eight Epithelial Features on VGH σ[σ[pixel intensity]/Mean[intensity in epi. nuclei]] Min[elliptic fit of epi. cont. regions] Sum of unclassified objects σ(|epi. cytoplasmic obj. and nuclear obj.|) σ(Max[blue pixel atypical epi nuclei]) Mean[border between epi. cytoplasmic obj.] Max[|atypical epi. nuclei|] Max[Min[green pixel intensity in epi. regions]] Alive Log-Rank P = 0.02 Low-Risk High-Risk Years A. H. Beck, A. R. Sangoi, S. Leung, R. J. Marinelli, T. O. Nielsen, M. J. van de Vijver, R. B. West, M. van de Rijn, D. Koller. "Systematic analysis of breast cancer morphology uncovers stromal features associated with survival." Science translational medicine 3.108 (2011) 18 Three New Stromal Features for Cancer Prognosis Stromal tissue provides structural support and serves as a matrix in which other cells are embedded. Intensity difference between stromal region and neighbors Presence of stromal regions without nuclei Three Stromal Features on VGH Low Risk Log-Rank P = 0.004 Alive High Risk Low Risk High Risk Avg. relative border of stromal spindle nuclei to stromal round nuclei Low-Risk High-Risk Years A. H. Beck, A. R. Sangoi, S. Leung, R. J. Marinelli, T. O. Nielsen, M. J. van de Vijver, R. B. West, M. van de Rijn, D. Koller. "Systematic analysis of breast cancer morphology uncovers stromal features associated with survival." Science translational medicine 3.108 (2011) 19 The Future of Histopathology? Real time digitally assisted analysis Fully digital analysis New prognostic/diagnostic feature sets advance basic research 20 Summary § Machine learning has seen renewed interest in recent years § Impressive progress in software development and classification accuracy § Less impressive progress in hardware implementation of the most promising networks (convolutional neural networks) › But we are hoping to fix this… § Given these trajectories, we hope to take steps toward fully digital histology › Initial focus on algorithms, software › Potential end-game solution is a low-cost, power efficient custom machine/computer 21