Successfully reported this slideshow.
Your SlideShare is downloading. ×

DevFest19 - Early Diagnosis of Chronic Diseases by Smartphone AI

Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Upcoming SlideShare
Ai in healthcare
Ai in healthcare
Loading in …3
×

Check these out next

1 of 55 Ad

DevFest19 - Early Diagnosis of Chronic Diseases by Smartphone AI

Download to read offline

Session by Sabyasachi Mukhopadhyay
Kolkata Lead, Facebook Developer Circle
GDE in ML
Intel Software Innovator
Visiting Faculty, SCIT Pune
Co-Founder & Chief Research Officer, Twelit MedTech Pvt. Ltd

Session by Sabyasachi Mukhopadhyay
Kolkata Lead, Facebook Developer Circle
GDE in ML
Intel Software Innovator
Visiting Faculty, SCIT Pune
Co-Founder & Chief Research Officer, Twelit MedTech Pvt. Ltd

Advertisement
Advertisement

More Related Content

Slideshows for you (18)

Similar to DevFest19 - Early Diagnosis of Chronic Diseases by Smartphone AI (20)

Advertisement

More from Gaurav Kheterpal (11)

Recently uploaded (20)

Advertisement

DevFest19 - Early Diagnosis of Chronic Diseases by Smartphone AI

  1. 1. Sabyasachi Mukhopadhyay Kolkata Lead, Facebook Developer Circle GDE in ML Intel Software Innovator Visiting Faculty, SCIT Pune Co-Founder & Chief Research Officer, Twelit MedTech Pvt. Ltd Early Diagnosis of Chronic Diseases by Smartphone AI
  2. 2. What problem should we solve? Use the process below to Identify, Articulate and Probe new Problems worth solving
  3. 3. Goal: What is a problem worth solving? Pick a theme e.g. ‘Digital Tech in Ag’ 1. Resource Map A. People and Orgs B. Users and beneficiaries C. Breakthroughs and risks 2. Problem Canvas 3. Solution Canvas 4. Findings Plot
  4. 4. 1. Resources Map A. People and Orgs B. Users and beneficiaries C.Breakthroughs and risks
  5. 5. 1A. Orgs/Lists • Orgs, Companies, Startups, Govt. dept • Data sets and information sources • Partners • People, VCs, Influentials • Known Published Challenges (e.g. UN, Gates, NSF, NIH, DBT/DST)
  6. 6. 1B. Users and beneficiaries • Graph of users and use cases • Influencers • Buyers/sponsors • Partners • Where do we need research? • Speak with experts!
  7. 7. 1C. Breakthroughs, benefits and risks • Breakthroughs, Parallel tech, Trends • Payoffs, results, benefits  IMPACT • Friction – regulation, biases, etc. • Financial risks • Execution risks/ limiting factors (e.g., time, distance, lack of existing infrastructure)
  8. 8. 2. Problem Canvas
  9. 9. 3. Solution Canvas
  10. 10. 4. Findings Plot • Take a guess on two major factors that matter to you (usually cost and time).Cost is a proxy for complexity etc. Time can be proxy for Engg time, regulation, market penetration etc.
  11. 11. MILESTONES Map of users, resource requirements, dependencies, risks and workarounds, steps, decision points ----------
  12. 12. Guides
  13. 13. Heilmeier questions* 1. What are you trying to do? Articulate objectives without using jargon. 2. How is it done today? What are the limits of current practice? 3. What’s new in your approach? Why do you think it will be successful? 4. Who cares? 5. If you’re successful, what difference will it make? 6. What are the risks and payoffs? 7. How much will it cost? 8. How long will it take? 9. What are the midterm and final “exams” to check for success? * 9 questions to develop a meaningful project
  14. 14. LEAN CANVAS
  15. 15. Let’s Dive In
  16. 16. Introduction  Optical Methods for tissue characterization Light scattering form Biological tissue Elastic Scattering spectra from biological tissue Morphological information: Shape and size, refractive index, intra-cellular organelles etc. Micro-optical property: Fractal parameters, disorder strength Inelastic Scattering: Fluorescence and Raman Spectroscopy Biochemical Information
  17. 17. Present interest: Light scattering based inverse analysis models – quantification of sub- diffractional changes in tissue refractive index fluctuations.  Monofractal approximation may not be realistic in biological tissue having wide range of dimensions and the complex nature of the spatial correlations  Spatial distribution of refractive index (RI) in biological tissue observed to exhibit statistical self-similarity  Analyzed using fractal-Born approximation --------- Phys. Rev. Lett. 97, 138102 (2006). Multifractal? Multiple scaling exponents Multi-resolution analysis  Scale invariant, power law dependence, single scaling exponent
  18. 18. • Fractal / Self-similarity are process / geometry, which repeat its pattern in different scale of space or time. • Fractal dimension: measure the space filling capacity in geometry. Fractals (self-similar) and Multifractals ( )P      2 EH D    Follows power law: ;  being the spatial frequency. Fractal properties quantified by a single scaling exponent, , where 𝑫 𝑬 = Euclidian Dimension, H = Hurst exponent; 0 < H <1.  Fractal dimension: 𝐷𝑓 = 𝐷 𝐸 − 𝐻 + 1.
  19. 19. • Observations- • Power law scaling in wide range of spatial frequency with background randomness-complexity of RI fluctuations • Multiple scaling exponent is observed- Fourier analysis is not adequate--- ---- signature of multifractality? DIC image: Connective Tissue Unfolding the image to 1D Spatial Refractive Index (RI) variation Fourier analysis Observing multifractality in spatial refractive index fluctuations  Required a better multiresolution analysis to characterize such Complex self-similar behavior.  Multifractal Detrended Fluctuation Analysis (MFDFA) is one of the state of the art multi-resolution analysis can be used.
  20. 20. Data   NixxiY i k k ...,1,)( 1   )/int( sNNs     s i iyisY s sF 1 22 )(])1[( 1 ),(    q N q s q s sF N sF /1 2 1 2/2 ),( 2 1 )(          )( )( qh q ssF  Multifractal Detrended Fluctuation Analysis (MFDFA) Profile Y(i) Divide Y(i) Local Fluctuation Fluctuation Function Log-Log Plot of s vs. Fq(s) h(q)= log[Fq(s)] log(𝑠) τ 𝑞 = 𝑞ℎ 𝑞 − 1 α = τ′ 𝑞 f(α) = qα - τ 𝑞 H. Eugene Stanley et al, Physica A 316 (2002) 87 – 114.  Width of the singularity spectrum (f(α)): 𝝈  N - length of the fluctuation series  s - length of each segments  Order of moments q )(sFq Negative q capture small fluctuations  Classical scaling exponent, 𝜏(q)  Hurst exponent, h(q)
  21. 21. Confirmation of multifractality: MFDFA analysis on DIC image Multifractality confirmed: ---- Significant variation of slope with moment q - 𝒍𝒐𝒈[Fq(s)] vs. log(s)
  22. 22. Quantification of Multifractality in Different Grades of precancerous Tissue  Differences more prominent for negative q-values h(q=2)  at Higher grade  Width of singularity spectrum (σ)  at Higher grade Small scale (sub-m) fluctuations! Stronger multifractality Finer index fluctuations (roughness / heterogeneity)
  23. 23. Tissue Region Hurst exponent ( mean h(q=2) ± standard deviation) Width of singularity spectra (mean 𝛔 ± standard deviation) Gr-I Gr-II Gr-III Gr-I Gr-II Gr-III Connective Tissue 0.54 ± 0.03 0.50 ± 0.04 0.36 ± 0.08 0.60 ± 0.10 0.68 ± 0.13 0.88 ± 0.07 Spatial refractive index variation in biological tissue exhibits multifractality direct evidence (MFDFA on DIC image)  Need of the hour: Extracting multifractal information from Light scattering -------- clinically amenable and in-situ deployable Key results MFDFA results on different pathology grades SPIE Proceedings, 2014
  24. 24. Scheme of extracting multifractality from light scattering spectra Elastic scattering spectra from tissue Fourier domain inverse pre- processing Subjected to MFDFA Strength of Multifractality = 𝛔; Hurst exponent: h(q=2) 𝜼/ 𝝆 : representative index inhomogeneity distribution in spatial scale ρ
  25. 25. Extraction of multifractality from tissue light scattering spectra Scattering spectra from tissue Experimental light scattering spectroscopic system Multiple power law exponent --- signature of multifractality? Wave length variation [I()]; Fixed 
  26. 26. Sample Preparation Biopsied cervical precancer tissue slices (The histopathologically characterized grade I, grade II, grade III precancer tissue) and normal tissues were collected from Ganesh Shankar Vidyarthi Memorial (GSVM) Medical College, Kanpur, India (age of patients between 35 – 60 years; ntotal = 35, with ngrade I = 14, ngrade II = 6, ngrade III = 9; four biopsies from the normal counterparts, nnormal = 6). The Standardized histological preparation of the excised tissues involving fixation, dehydration, imbedding in wax, sectioning under a rotary microtome with thickness ∼5 μm, lateral dimension ∼4 mm × 6 mm, is followed by performing subsequent de-waxing. The consent for the use of all the intact tissue (human cervix with cancer and normal) samples in our study was obtained from the Ethical Committee, G. S. V. M. Medical College and Hospital, Kanpur, India. The sample preparation methods follow approved guidelines in our study.
  27. 27. Experimental System The spatial distribution of tissue refractive index was recorded by a differential interference1 contrast (DIC) microscope (Olympus IX-81, USA). At a magnification of 60X, these DIC2 Images were recorded by a CCD camera (ORCA-ERG, Hamamatsu, 1344 X 1024 pixels pixel3 dimension 6.45 m). The elastic scattering spectra from the multiple sites of the biopsied tissue4 sections were recorded by the angle resolved spectral light scattering measurements (Figure 1).5 In brief, light emitted from a Xe-lamp (HPX-2000, Ocean Optics, USA) was collimated by a6 combination of lenses and illuminated the tissue sample at the centre of a goniometric7 arrangement (spot size ~1-mm-diameter). The collimated scattered light from sample was8 focused into a collecting fiber probe coupled to a spectrometer (USB4000FL, Ocean Optics,9 USA) for wavelength resolved signal detection. The recordings of spectra were performed (360-10 800nm) with a spectral resolution of 2.05 nm, where the angular range was kept at 10 150 11 with an interval of10. For the inverse multifractal study the spectra were recorded at12 backscattering angle  = 150o 13
  28. 28. Schematic of the spectral light scattering measurement. Xe lamp: excitation source; A: aperture; L1: collimating lens;L2: illuminating lens; L3 and L4: collecting lenses; f1: Focal length of collimating lens L1; f2: focal length of illuminating lens L2 ; f3 and f4: Focal lengths of collecting lenses L3 and L4 respectively.
  29. 29. Extraction and quantification of tissue multifractal parameters  Light scattering intensity from Grade I tissue  Detrended refractive index Fluctuations extracted from Fourier domain preprocessed signal  Multifractality confirmed:- Significant variation of slopes with moment q- 𝒍𝒐𝒈[Fq(s)] vs. log(s)  Prominent variation in negative q-values Small scale (sub-m) fluctuations
  30. 30. Summary of results on tissue  Consistency of the multifractal trends  statistical significance of the observed differences in the multifractal parameters No. of tissue samples studied: Healthy-20,Grade I -15, Grade II- 15 and Grade III -25  Hurst exponent, h(q=2)  --- implies overall reduction in the depth correlation of refractive index  Strength of multifractality, σ -- indicates the increased of roughness / inhomogeneity of depth resolved index variations  Possible reasons: fibrous network in connective tissue gets fragmented due to the shortening / breaking of the building blocks, the collagen fibers / micro-fibril with the progression of cancer / precancer. J. Opt. 18(12), 125301 (2016) Normal Grade-I Grade-II Grade-III Hurst Exponent(h(q=2)) 0.630.02 0.560.05 0.480.03 0.410.04 Singularity Spectrum Width (∆α) 0.860.01 0.900.03 0.960.04 0.990.01
  31. 31. Our initial dataset consists of 75 samples. We have used Monte Carlo cross-validation, where we randomly split the dataset into training and testing dataset. This process has been repeated 100 times. The size of training and testing dataset varies for each split; we just ensured that a minimum (2) no of samples have been included in both training and testing dataset for each split. For each such split, the model has been fit to the training data, and predictive accuracy has been assessed using just the validation data (testing data). Then the results of the testing data have been averaged over the splits. The advantage of this method is that the proportion of the training/validation split is not dependent on the number of iterations and the predictive accuracy is more or less independent of the samples used for training dataset. We additionally have used 9 unknown samples taken at different time than the dataset and the prediction done by our model has been compared by manual verification. Results of the unknown samples also have been averaged and presented in the results. The above process has been repeated for SVM and HMM.
  32. 32. Support vector machine (SVM) based multiclass classification on extracted multifractal parameters from tissue samples.
  33. 33. SVM based tissue classification based on light scattering –derived multifractal tissue optical properties. The horizontal surfaces of these figures have been sectioned into 4  4 rectangles. The accurate and inaccurate prediction of each stages (normal, grade I, grade II, grade III) have been represented by diagonal and off diagonal rectangles respectively. Normal Grade I Grade II Grade III 0 10 20 30 40 50 60 70 80 90 100 Normal Grade I Grade II Grade III 98.5 14.29 10 0 1.5 57.14 35 0 0 28.57 55 0 0 0 0 100 Normal Grade I Grade II Grade III
  34. 34. The flowchart of HMM based model on multifractal tissue optical properties derived from light scattering spectra.
  35. 35. HMM based tissue classification based on light scattering –derived multifractal tissue optical properties. The horizontal surfaces of these figures have been sectioned into 4  4 rectangles. The accurate and inaccurate prediction of each stages (normal, grade I, grade II, grade III) have been represented by diagonal and off diagonal rectangles respectively.
  36. 36. The results demonstrate that binary classification between normal and cancerous tissues (grade III) are very good both in SVM and HMM. Meanwhile in multiclass classification cases, when precancerous grades (grade I, grade II) are to be classified along with normal and precancerous tissues (grade III), abstract parameters achieved using HMM performs better than the SVM. The presence of noise in the obtained signal damages the SVM performance as SVM clearly classifies based on the kernel formed after considering all the multifractal parameters. While in the case of HMM, the Markov model finds abstract parameters by controlling the actual multifractal parameters and produces prediction based on the derived abstract parameters. As a consequence HMM avoids the noise added to the signal and able to produce better multiclass classification results than SVM.
  37. 37. Probing depth resolved multifractality in human retina from Optical Coherence Tomography (OCT) image  OCT can extract depth resolved index variation  Analyzed via MFDFA to extract alteration of multifractality due to progress of diabetic macular edema (DME) and age related macular degeneracy. • Cropped different layers of retina  Acquired in-vivo retinal OCT  Extracted depth resolved optical index variations of different layers
  38. 38. Variation of depth resolved optical index from OCT Images of in -vivo human retina  ILM: Inner limiting membrane  NFL: Nerve fiber layer  GCL: Ganglion cell layer  IPL: Inner plexiform layer  INL: Inner nuclear layer  OPL: Outer plexiform layer  ONL: Outer nuclear layer  ELM: External limiting membrane  IS/OS: Inner and outer photoreceptor  RPE: Retinal pigment epithelium  OPR: Outer segment PR/RPE complex Diabetic macular edema
  39. 39. Quantification of multifractality in depth resolved optical index variation in in-vivo human retinal layers Prominent difference in negative q indicates DME related index alters mainly in small length scale Analysis performed on Outer plexiform layer(OPL)  h(q=2)  in OPL of diabetic macular edema (DME): indicates reduction of optical index correlation in progress of DME  σ  in OPL of diabetic macular edema (DME): indicates increase of strength of multifractality or heterogeneity in layers with progress of DME variation of 𝒍𝒐𝒈[Fq(l)] vs. log(l) slope variation of h(q): multifractality
  40. 40. Retinal Layers h(q=2) σ Healthy AMD Healthy AMD NFL 0.87 ± 0.04 0.66 ± 0.01 0.62 ± 0.08 0.93 ± 0.09 GCL 0.81 ± 0.03 0.61 ± 0.02 0.85 ± 0.02 1.57 ± 0.07 IPL 0.72 ± 0.02 0.66 ± 0.03 0.77 ± 0.05 1.23 ± 0.08 OPL 0.62 ± 0.01 0.53 ± 0.02 0.65 ± 0.09 1.00 ± 0.10 ONL 0.61 ± 0.03 0.51 ± 0.02 1.02 ± 0.07 1.31 ± 0.06 Choroid 0.76 ± 0.04 0.57 ± 0.03 0.82 ± 0.08 1.27 ± 0.12 Results of multifractality in healthy and age related macular degeneracy (AMD) in in-vivo human retinal layers  Hurst exponent, h(q=2)  --- implies overall reduction in the depth correlation of refractive index  Strength of multifractality,σ  -- indicates the increased of roughness / inhomogeneity of depth resolved index variations J. Biomed. Opt. 21(9), 096004 (2016)
  41. 41. [1] N.Das, S.Mukhopadhyay, N.Ghosh, J.Chhablani, A.Richhariya, K.D.Rao, N. K.Sahoo, “Investigation of Alterations in multifractality in Optical Coherence Tomographic Images of In Vivo Human Retina”; Journal Of Biomedical Optics 21(9), 096004 (Sep 09, 2016). [2] S.Mukhopadhyay, N.Das, I.Kurmi, A.Pradhan, N.Ghosh, P.K.Panigrahi, “Tissue multifractality and hidden Markov model based integrated framework for optimum precancer dectection”, Journal Of Biomedical Optics 22(10), 105005 (Oct 19, 2017). [3] S.Mukhopadhyay et al., “Recurrence Quantifications as Potential Bio-markers for Diagnosis of Pre-Cancer”, SPIE Proceeding, SPIE Photonics West, 2017, USA. [4] S.Mukhopadhyay et al., “Optical Diagnosis of Cervical Cancer by Intrinsic Mode Functions”, SPIE Proceeding, SPIE Photonics West, 2017, USA. [5] S.Pratiher, S.Mukhopadhyay, R.Barman, S.Pratiher, A.Pradhan, N.Ghosh, P.K.Panigrahi, “Optical Diagnosis of Cervical Cancer by Higher Order Spectra and Boosting”, SPIE Proceeding, SPIE Photonics West, 2017, USA. [6] S.Mukhopadhyay et al., “Optical diagnosis of colon and cervical cancer by support vector machine”, SPIE Proceeding, SPIE Photonics Europe, Belgium, 2016, Europe. [7] S.Mukhopadhyay et al., “S-TRANSFORM BASED FLUCTUATION ANALYSIS- A METHOD FOR PRECANCER DETECTION”, IEEE Conference Proceeding, Microcom-2016, India. [8] S.Mukhopadhyay et al., “Wavelet and multi-fractal based analysis on DIC images in epithelium region to detect and diagnose the cancer progress among different grades of tissues”, SPIE Proceeding, SPIE Photonics Europe, Belgium, 2014, Europe. [9]S.Mukhopadhyay et al., “Pre-cancer Detection by Wavelet Transform and Multi-fractality in various grades of DIC Stromal Images”, SPIE Proceeding, SPIE Photonics West, 2014, USA. [10] S.Mukhopadhyay et.al., “Efficacy of hidden Markov model over support vector machine on multiclass classification of healthy and cancerous cervical tissues”, SPIE Proceeding, SPIE Photonics West, 2018, USA. [11]S.Mukhopadhyay et al., “A two-stage framework for DIC image denoising and Gabor based GLCM feature extraction for pre-cancer diagnosis”, SPIE Proceeding, SPIE Photonics West, 2018, USA. My Publications Related To Early Stage Disease Detection
  42. 42. 9 Impact on BOP Level Pain free Early Stage Cancer Diagnosis and Low Cost Portable Device First Solution Generate report in a few minutes where biopsy examination takes several days Second Solution Automatic Solutions and Highly Efficient Third Solution Stay Away From Tobacco : Be Safe in the Sun : Eat Healthy and Get Active
  43. 43. Smart Phone based Cancer & DME Solutions
  44. 44. Unique Selling Proposition 100 Times Cheaper than Biopsy and Accuracy >95% Portable & Lightweight and 80 Times Faster than Biopsy Replacing Painful and Cost Effective Biopsy
  45. 45. Business Model B2B Product for Ophthalmologists DeviceSaaS Rs100/Scan One Time Rs20000
  46. 46. Growth Strategiesowth Strategy FutureCurrent Ongoing Clinical Trail in Hospitals Doctors Collaboration 50- 50 Revenue Sharing Model
  47. 47. Competitive Landscape of Our Innovation Twelit Low Cost Accuracy >95% Smart Device Data Driven Cloud Based
  48. 48. Media Reports on Published Research: New technique for early detection of human eye diseases Televisions: [1]Zee News: http://zeenews.india.com/health/this-new-technique-can-detect-human-eye- diseases-early-1962133 [2]Odisha TV: https://odishatv.in/technology/new-technique-for-early-detection-of-human- eye-diseases-184520/ [3]News World India: https://newsworldindia.in/lifestyle/now-early-detection-of-human- eye-diseases-is-possible-know-how/240282/ [4]NewsX: http://www.newsx.com/health-and-science/50839-new-technique-for-early- detection-of-human-eye-diseases Global News Site: Yahoo News: https://in.news.yahoo.com/technique-early-detection-human-eye-diseases- 081403358.html Science Magazine: Nature India: http://www.natureasia.com/en/nindia/article/10.1038/nindia.2016.169
  49. 49. Media Reports on Published Research: New technique for early detection of human eye diseases Newspapers [1]The Hindu: http://www.thehindu.com/sci-tech/science/Indian- scientists%E2%80%99-novel-approach-to-diagnose-retinal- diseases/article17004895.ece#comments [2]Business Standard: http://www.business-standard.com/article/news-ians/new- technique-for-early-detection-of-human-eye-diseases-116122700377_1.html [3]The Indian Xpress: https://theindianxpress.com/20056/ [4]The Economic Times: http://economictimes.indiatimes.com/news/science/new- technique-for-early-detection-of-human-eye- diseases/articleshow/56198412.cms?from=mdr
  50. 50. 24 24. Cancer Research Success Newspapers & Research Magazines: [1]The Hindu:http://www.thehindu.com/sci-tech/science/diagnosing-early-stage- cervical-cancer-using-artificial-intelligence/article22267117.ece [2]Business Standard:http://www.business-standard.com/article/news- ians/computer-based-optical-method-detects-early-stage-cervical-cancer- 117122800553_1.html [3]The Indian Express: http://indianexpress.com/article/lifestyle/health/computer- based-optical-method-detects-early-stage-cervical-cancer-5002420/ [4]Analytics India Magazine: https://analyticsindiamag.com/ai-cancer-detection-india- light-scatter-algorithm/ [5] Nature India :https://www.natureasia.com/en/nindia/article/10.1038/nindia.2017.148 Stay Away From Tobacco : Be Safe in the Sun : Eat Healthy and Get Active AI based early stage cancer detection
  51. 51. IBPS,Banking,IAS,UPSC Exams Current Affair Syllabus [1] http://bankexamportal.com/daily-current-affairs/24- december-2017 [2]http://iasexamportal.com/civilservices/daily-current- affairs/24-12-2017
  52. 52. 25 Award On AI Based Early Stage Disease Detection My futuristic research on ‘AI Based Early Stage Disease Detection’ was nominated among top 10 finalists in Global IMT X-Prize international competition in USA. Dr.Barmak Heshmat, Research Scientist, MIT was one of the organizers of IMT X-Prize global competition. The panel comprised of eminent scientists and artists from MIT, NASA, Marvel Studios (Franchises of block buster Hollywood movies like X-Men, Spider Men). The competition theme was to produce futuristic ideas for 'AI (Artificial Intelligence) for Good'. I was among top 10 finalists but ended as a runner up.
  53. 53. [1] We are currently expanding our investigations towards in-vivo deployment of this integrated approach for precancer detection using tissue light scattering spectra. [2] Industrial implications are going on at LVPEI, Hyderabad for early stage DME and diabetic retinopathy detection of eyes. Ongoing and Future Studies
  54. 54. @SM2017Official @sabyasachi_unique https://www.linkedin.com/in/sabyasachi-mukhopadhyay-303a1027/ @sabyasachi_mukhopadhyay Join me:
  55. 55. THANK YOU For your time & we’ll see you soon

×