Mining Medical Images R. Bharat Rao Glenn Fung Balaji Krishnapuram Jinbo Bi Murat Dundar Vikas Raykar Shipeng Yu Sriram Krishnan Xiang Zhou Arun Krishnan Marcos Salganicoff Luca Bogoni Matthias Wolf Anna Jerebko Jonathan Stoeckel Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Outline of the talk Mining medical images Computer aided diagnosis (CAD) Key data mining challenges Clinical impact Lessons learnt Several thousand units of the products described in this paper have been commercially deployed in hospitals around the world since 2004 Page 2 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Medical Imaging 1895 X-ray used for broken bones, locating foreign objects 1970 Computed tomography (CT) 3-D imaging As resolution increased in-vivo imaging is widely used to locate medical abnormalities for diagnosis and surgery planning Digital Mammogram CT Scan Page 3 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Mining medical imaging data Increased resolution has resulted in Data Overload Increased total study time Increase in data does not always translate to improved diagnosis Automatically extract the actionable information from the imaging data in order to ensure improvement in patient care simultaneous reduction in total study time Clinically relevant information Raw imaging data Knowledge based data-mining algorithms Computer aided diagnosis/detection CAD Page 4 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Computer-aided diagnosis/detection (CAD) CAD technologies support the physician by drawing attention to structures in the image that may require further review. Used as a second reader Improves the detection performance of a radiologist Reduces mistakes related to misinterpretation The principal value of CAD is determined by carefully measuring the incremental value of CAD in normal clinical practice Page 5 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Lung CAD Identify suspicious regions called nodules (which are known to be precursors of cancer) in CT scans of the lung. Page 6 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Colon PEV Polyp Enhanced Viewer Identify suspicious regions called polyps in CT scans of the colon. Page 7 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Mammo CAD Identify abnormal masses/calcifications in digital mammograms. PECAD and MammoCAD are only sold outside the US. Page 8 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. PE CAD Pulmonary Embolism (PE) is a sudden blockage in a pulmonary artery caused by an embolus that is formed in one part of the body and travels to the lungs in the bloodstream through the heart. PECAD and MammoCAD are only sold outside the US. Page 9 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. CAD Goal is to detect potentially malignant nodules (lung) polyps (colon) lesions (breast) Pulmonary emboli (lung) in medical images like CT scans, X-ray, MRI, etc. Early detection provides the best prognosis Page 10 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Typical CAD architecture Potential candidates Image [ X-ray | CT scan | MRI ] Candidate Generation > 90% sensitivity 60-300 FP/image Feature Computation Lesion Classification Location of lesions > 80% sensitivity 2-5 FP/image Focus of the current talk Page 11 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Key Data Mining Challenges High accuracy 2-5 FP/image sensitivity > 80% 1. The breakdown of assumptions 2. Highly unbalanced data 3. Feature computation cost 4. Incorporating domain knowledge 5. No objective ground truth Page 12 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. The breakdown of assumptions region on a mammogram lesion not a lesion Traditional classification algorithms Neural networks Support Vector Machines Logistic Regression …. Make two key assumptions (1) Training samples are independent (2) Maximize classification accuracy over all candidates Page 13 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Often violated in CAD Violation 1: Training examples are correlated Candidate generation produces a lot of spatially adjacent candidates. Hence there are high level of correlations among candidates. Also correlations exist across different images/detector type/hospitals. Page 14 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Violation 2: Candidate level accuracy is not important Most algorithms maximize classification accuracy. Try to classify every candidate correctly. Several candidates from the CG point to the same lesion in the breast. Lesion is detected if at least one of them is detected. It is fine if we miss adjacent overlapping candidates. Hence CAD system accuracy is measured in terms of per lesion/image/patient sensitivity. So why not optimize the performance metric we use to evaluate our system? Page 15 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Solution 1: Multiple Instance Learning Fung, et al. 2006, Bi, et al. 2007, Raykar et al. 2008, Krishnapuram, et al. 2008, How do we acquire labels ? Candidates which overlap with the radiologist mark is a positive. Rest are negative. Single Instance Learning 1 Multiple Instance Learning Positive Bag 1 1 0 0 0 0 0 0 0 0 Classify at-least one candidate correctly Classify every candidate correctly We have modified SVM and logistic regression for multiple instance learning Page 16 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Simple Illustration Accounts for correlation during training Single instance learning: Multiple instance learning: Reject as many negative candidates as possible. Reject as many negative candidates as possible. Detect as many positives as possible. Detect at-least one candidate in a positive bag. Multiple Instance Learning Single Instance Learning Page 17 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Solution 2: Batch Classification Vural et al., 2009 Accounts for correlation during testing Change the decision boundary during test time. Page 18 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Skewed data and expensive features 1. Highly unbalanced class distribution (less than 1% are abnormal) 2. Huge number of experimentally engineered features 3. Lot of them are irrelevant and redundant. 4. Feature computation is expensive 5. Stringent run-time requirements 1. Feature selection/Sparse classifiers 2. Cascaded classification architecture Page 19 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Cascaded classification architecture Bi, et al. 2006 Page 20 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Novel AND-OR training of cascades Dundar and Bi 2007 Page 21 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Incorporating domain knowledge We know that lesions have different shapes/sizes/appearance Page 22 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Gated Classification architecture Page 23 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Incorporating domain knowledge Dundar et al. 2007 Exploit different sub-classes of False Positives Page 24 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Subjective Ground truth Raykar et al. 2009 Each radiologist is asked to annotate whether a lesion is malignant (1) or not (0). Lesion ID Radiologist 1 Radiologist 2 Radiologist 3 Radiologist 4 Unknown 12 0 0 0 0 x 32 0 1 0 0 x 10 1 1 1 1 x 11 0 0 1 1 x 24 0 1 1 1 x 23 0 0 1 0 x 40 0 1 1 0 x We have proposed an EM algorithm to simultaneously learn the ground truth and the classifier. In practice there is a substantial amount of disagreement. Page 25 Truth We have no knowledge of the actual golden ground truth. Getting absolute ground truth (e.g. biopsy) can be expensive. Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Key Data Mining Challenges Challenge Solutions 1. Training/testing data is correlated Multiple instance learning batch classification 2. Evaluation metric is CAD specific Multiple instance learning 3. Highly unbalanced data Cascaded classifiers 4. Feature computation cost Cascaded classifiers Feature selection methods 5. Incorporating domain knowledge Gated classifiers Polyhedral classifiers 6. No objective ground truth EM algorithm Page 26 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Clinical Impact 1. How much can a radiologist benefit by using the CAD software ? 2. CAD is mostly deployed in second reader mode. 3. Measure the improvement in performance of a radiologist with CAD. 4. Several independent clinical studies/trials have been conducted by our collaborators worldwide. Page 27 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Lung CAD 1. FDA clinical validation study with17 radiologists,196 cases from 4 hospitals. Average reader AUC increased by 0.048 (p<0.001) because of CAD. 2. Recent study at NYU by Godoy et al. 2008 Sensitivity without CAD Sensitivity with CAD Increase in sensitivity 56.2 % 79.2 % 66.0 % 89.8 % 9.8 % 10.6 % Reader 1 Reader 2 3. New prototype also helps detect different kinds of nodules. Solid Nodules Part-solid Nodules Ground Glass Opacities Page 28 Mean sensitivity without CAD Mean sensitivity with CAD Increase in sensitivity 60% 80% 75% 85% 95% 86% 15 % 15% 11% Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Colon PEV Colon PEV (Polyp Enhanced Viewer) was evaluated by Baker, et al. 2007 Study with seven less-experienced readers Without PEV average sensitivity was 0.810 With PEV average sensitivity was 0.908 A 9.8% increase in average sensitivity (p=0.0152). Page 29 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. PE CAD Das et al. 2008 conducted a study with 43 patients to asses the sensitivity of detection of pulmonary embolism. . Sensitivity without CAD Sensitivity with CAD Increase sensitivity Reader 1 87% 98% 11% Reader 2 82% 93% 11% Reader 3 77% 92% 15% Page 30 in Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Key data mining lessons 1. True measure of impact is how much does CAD help the radiologists. 2. Design algorithms to optimize the metric you care about 3. Careful analysis of the assumptions behind off-the-shelf data-mining algorithms. In CAD most of these assumptions break down. Need to design new methods. 4. Domain knowledge is very important. Collaboration with radiologists is crucial in eliciting the domain knowledge. Page 31 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Conclusions 1. Radiologists have access to orders of magnitude more data for diagnosing various cancers. 2. Difficult and time-consuming to identify key clinical findings. 3. We described the data-mining challenges in a commercially deployed CAD software. 4. Use of CAD as second reader improves radiologist's detection performance. 5. Key opportunity for data mining technologies to impact patient care worldwide. Page 32 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved. Acknowledgements Dr. D. Naidich, MD, of New York University Dr. M. E. Baker, MD, of the Cleveland Clinic Foundation Dr. M. Das, MD, of the University of Aachen Dr. U. J. Schoepf, MD, of the Medical University of South Carolina Dr. Peter Herzog, MD, of Klinikum Grossharden, Munich. Alok Gupta, Ph.D., Ingo Schmuecking, MD, Harald Steck, Ph.D., Stefan Niculescu, Ph.D., Romer Rosales, Ph.D., Sangmin Park, Ph.D., Gerardo Valadez Ph.D. Maleeha Qazi, and the entire SISL team. Page 33 Copyright © 2009 Siemens Medical Solutions USA, Inc. All rights reserved.