Approaches to Characterization of Indeterminate Lung Nodules Denise
Approaches to Characterization of Indeterminate Lung Nodules Denise R. Aberle, MD Professor of Radiology and Bioengineering David Geffen School of Medicine at UCLA Overview Rationale and differences in approaches Semantic feature analysis Hand-crafted, quantitative features Deep learning Semantic feature analysis Visual characterization of object(s) using commonly understood terms & definitions Time tested Requires minimal software (DICOM viewer, calipers, ROI)
Agnostic to image acquisition/reconstruction parameters Highly labor intensive Not quantitative Moderate to high inter-reader variability Semantic features: The Brock model (screening) Developed and validated in separate screening cohorts Developed: PanCan | 1871 persons had 7008 nodules Validated: BCCA | 1090 persons had 5021 nodules Average of 5 to 7 nodules per person Combined clinical and imaging variables Both parsimonious & full models | AUC > 0.90
Nodules of all sizes (high number nodules < 4 mm) Brock Model (Parsimonious & Full) 9 Predictor Variables Parsimonious Model OR (95% CI) P value 1.91 (1.19-3.07) 0.008 Non-linear < 0.001 1.82 (1.12-2.98) 0.02 2.54 (1.45-4.43) 0.001 Age (Centered at 62; per year) Sex (Female vs. Male) FH (Yes vs. No) Emphysema (Yes vs. No) Nodule size (Centered at 4 mm) Nodule consistency:
GGN PSN Solid Lobe: Upper vs. other Nodule count per scan Spiculation (Yes vs. No) In test set: Parsimonious model AUC = 0.960 (0.927-0.980) M. NEJM 2013. McWilliams A etMcWilliams al. NEJM 2013;369:910-919. Brock Model (Parsimonious & Full) 9 Predictor Variables Parsimonious Model OR (95% CI) P value Full Logistic Regression OR (95% CI) P value 1.03 (0.99-1.07) 0.16 1.82 (1.12-2.97) 0.02
FH (Yes vs. No) 1.34 (0.83-2.17) 0.23 Emphysema (Yes vs. No) 1.34 (0.78-2.33) 0.29 Non-linear < 0.001 GGN 0.88 (0.48-1.62) 0.68 PSN 1.46 (0.74-2.88) 0.28 Solid Reference Age (Centered at 62; per year)
Sex (Female vs. Male) 1.91 (1.19-3.07) Nodule size (Centered at 4 mm) Nodule consistency: Lobe: Upper vs. other Non-linear 1.82 (1.12-2.98) 0.008 < 0.001 0.02 Nodule count per scan Spiculation (Yes vs. No) 2.54 (1.45-4.43) 0.001 1.93 (1.14-3.27) 0.02 0.92 (0.85-1.00) 0.049
2.17 (1.16-4.05) 0.02 In test set: Parsimonious model AUC = 0.960 (0.927-0.980) | Full model AUC = 0.970 (0.945-0.986) M. NEJM 2013. McWilliams A etMcWilliams al. NEJM 2013;369:910-919. Diagnostic prediction in screening setting | NLST CT arm of NLST using Baseline positive screens (N = 11,128 nodules in 6726 participants) At nodule level (4-30 mm): Predict lung cancer in same lobe as nodule Logistic regression with 10-fold cross-validation AUC = 0.89 (0.87-09.03) using clinical + imaging Variable OR
95% CI P-value Age (per 1 year increase) 1.047 1.023 1.071 < 0.0001 Sex: female vs. male 1.277 1.000 - 1.631 0.0500 Family history of lung cancer 1.312 1.004 1.714 0.0469 Body mass index (per 1 unit) 0.972 0.947 0.997
0.0285 Smoking status (current vs. former) 1.226 0.960 - 1.564 0.1018 Pack-years (per 1 pack-year) 1.011 1.006 1.015 < 0.0001 Self-reported history of COPD 0.928 0.600 1.435 0.7368 Longest diameter (per 1 mm increment) 1.148 1.125 1.171 < 0.0001
Consistency: GGN vs. solid nodule 0.654 0.479 0.894 0.0077 Consistency: PSN vs. solid nodule 1.011 0.663 1.541 0.9605 Upper vs. middle lobe 3.815 2.451 5.937 < 0.0001 Upper vs. lower lobe 1.671 1.306 2.137 < 0.0001 Margins: spiculated/ill defined vs. smooth
2.796 2.132 3.667 < 0.0001 Nodule count per scan (per additional nodule) 0.897 0.823 0.978 0.0140 Nodule-specific features Radiologist level of suspicion for lung cancer (high/moderately high vs. low/no) 9.739 5.491 17.271 < 0.0001 Limitations of semantic analysis Time intensive Low to moderate reader agreement Solid vs. subsolid nodules | K = 0.61
Solid vs. PSN vs. GGN | K = 0.33 Margin characteristics Size variations depending upon perception of nodule boundary Yildririm A. Eur Congress Radiol; 03-2013. Van Riel S. RSNA 2013. Illustrated semantic atlas Textual definitions & representative images with consensus by thoracic radiologists General features: Size | anatomic location Consistency Margins (Conspicuity and Type) Shape features
Internal features External (Peri-nodule surround) features Consistency: Peri-cystic Margins Serrated/Spiculated Peri-nodule Stretching | Paracicatricial emphysema Progress Nearing completion of atlas Reader study Atlas naive radiologists characterize nodules pre- and post-training Goal: Determine degree of convergence on semantic labels
Determine robust semantic features Semantic analysis Make understandable the outputs of quantitative and machine learning features Publish atlas Quantitative hand-crafted features Quantitative | Reproducible on same analytical platform Requires dedicated analytical SW | Changes workflow Challenge of standardization across platforms Affected by segmentation
Hounsfield intensity value resampling (quantization) Mathematical formulae may be implemented differently Some quantitative features sensitive to technique Acquisition: kV | mAs Reconstruction: Spatial resolution (slice thickness) | Algorithm | Kernel Different groups are publishing independent analyses Limited multicenter efforts to compare results across common datasets Radiomics: Hand-crafted features Converts images to higher dimensional data Quantitative features are extracted from segmentations Frequency +1
-1 +1 Frequency -1 Gillies RJ. Radiology 2016; 278: 563-577/ Effects of dose, kernel, and reconstruction on histogram Homogeneous Water Phantom Lung nodule segmentation B10f B45f Diagnostic dose B70f 100% 30% 10% B10f B45f Low dose dose B70f
100% 30% 10% UPPER: Effect of three different kernels at diagnostic dose (17.1 mGy) LOWER: Effect of three different kernels at low dose (3.8 mGy). All images use filtered back projection (FBP). UPPER: Effect of three different doses using b45f kernel and FBP. LOWER: Effect of three different doses using I44f and IR. Lo P et al. Med Phys 2016; 43(8):4854-4865. Semantic features in screen-eligible patients Consecutive individuals undergoing percutaneous lung biopsy (80%) or being followed for indeterminate nodules (20%) Age 50 | ever smokers | 6-27 mm mean diameter | FU > 2 years N = 146 | 50 benign, 90 NSCLC, 6 well differentiated neuroendocrine Parsimonious logistic regression model Clinical & Semantic Features
OR 95% CI P-value Age (per 1 year increment) 1.061 1.007-1.118 0.027 Nodule size per mm 1.12 1.053-1.211 0.001 Consistency (sub-solid vs. solid) 3.579 1.162-11.020 0.026 Margin (non-smooth vs. smooth) 3.538
1.201-10.424 0.022 Radiologist suspicion 3.050 1.942-4.791 0.000 NPV = 86.84 | PPV = 84.26 | AUC = 0.90 (95% CI = 0.83-0.95) Quantitative feature analysis Consecutive individuals undergoing percutaneous lung biopsy (80%) or being followed for indeterminate nodules (20%) Segmentation: Automated, semi-automated, manual 1000 18 features (sequential forward floating selection with 20-fold cross-over sampling) 5 texture features using logistic regression with back feature selection Quantitative Features OR 95% CI P-value QMBLobness 40_skewness -1.997
NPV = 73.81 | PPV = 81.73 | AUC = 0.86 (95% CI = 0.77-0.91) Semantic and quantitative performance Deep learning using CNNs Nonlinear alternative to ad hoc descriptors for pattern recognition Input is the image Output is prediction of indolent vs. aggressive | benign vs. malignant Learns from the raw data itself: minimal preprocessing Not dependent on segmentation Relatively agnostic to image acquisition/reconstruction Requires LARGE data sets Several different architectures | transfer learning Data augmentation and scaling Data augmentation by generating
orthogonal images as inputs into CNN Scaling used to identify local pixel attenuations as well as larger context. Scales here are all 64 x 64 at 10, 20, 40 mm Ciompi F et al. Nature Scientific Reports 2017; 7:46479 | DOI: 10.1038/srep46479 CNN architecture Fully connected layer Ciompi F et al. Nature Scientific Reports 2017; 7:46479 | DOI: 10.1038/srep46479 Using semantic features for learning Train CNN with semantic labels to make output more intuitive to humans Assess classification performance for semantic labels & diagnostic prediction Comparison: 3D hierarchical semantic network AUC = 0.856 ( 0.026) 3D single CNN AUC = 0.847 ( 0.024) Mean difference = 0.005 (95% CI: 0.0051-0.0129); p = 0.009 Will Hsu, PhD
QOD 10/3/13 ... liberty, and property Petition of Rights Charles I (England) Limited the king's power King no longer imprison or punish any person by lawful judgment of his peers Challenged divine right of kings English Bill of Rights William...
Participants reported that VRS were slightly busier than FtF services . Image: Connaway, L.S. (2013) ... closed questions, some also from critical incident. Overall reference volume was reported as busy (11, 44%) or extremely busy (8, 32%). 40% reported that...
What is cultural competence? The integration and transformation of knowledge about individuals and groups of people into specific standards, policies, practices, and attitudes used in appropriate cultural settings to increase the quality of services, thereby producing better outcomes The ability...
Book Fifteen Athena summons Telemachus home and tells him how to avoid an ambush by the suitors. Meanwhile back on Ithaca, Odysseus listens while the swineherd Eumaeus recounts the story of his life. Eumaeus was the child of a prosperous...
The sum of these two types of movement gives the twelve degrees of freedom of a work piece in space. ... Example: lathe fixture , milling fixture etc. Drill jig: Drilling jig holds the work piece during drilling or allied...
Perceptual. Region. A region created by the interactions between a central node and surrounding locations. Functional Region. An area of near uniformity (homogeneity) in one or several characteristics. An area characterized by similarity or by cohesiveness that sets it apart...