Purpose: To predict the histopathologic subtypes with poor surgery prognosis in early stage lung adenocarcinomas using CT and PET radiomics.
Methods: We retrospectively enrolled 53 patients with stage I lung adenocarcinoma who underwent both diagnostic CT and 18F-fluorodeoxyglucose (FDG) PET/CT before complete surgical resection of the tumors. Tumor segmentation was manually contoured by a physician on both the diagnostic CT and the attenuation CT of PET/CT.A total of 170 radiomics features were extracted on both PET and CT images to design predictive models for two histopathologic endpoints: (1) tumors with solid or micropapillary predominant subtype (aggressiveness), and (2) tumors with micropapillary component more than 5% (MIP5). We used least absolute shrinkage and selection operator (LASSO) as a model building method coupled with a class separability feature selection (CSFS) method. For an unbiased model estimate, a 10-fold cross validation approach was used. The area under the curve (AUC) and prediction accuracy were employed to evaluate the performance of the model. P-values were computed using Wilcoxon rank-sum test.
Results: Of the 53 patients, 9 and 15 had tumors with aggressiveness and MIP5, respectively. For both endpoints, LASSO models with two PET radiomics features achieved the best performance. For aggressiveness, the LASSO model with PET Cluster Shade and PET 2D Variance resulted in 77.6±2.3% accuracy and 0.71±0.02 AUC (P = 0.011). For MIP5, the LASSO model with PET Eccentricity and PET Cluster Shade resulted in 69.6±3.1% accuracy and 0.68±0.04 AUC (P=0.014). The PET Cluster Shade was commonly selected in both models. Cluster shade is a texture feature that measures the skewness of the co-occurrence matrix. Higher PET cluster shade predicted that the tumor was more aggressive and more likely MIP5.
Conclusion: We showed that PET/CT radiomics features can predict tumor aggressiveness.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the National Cancer Institute Grants R01CA172638.
The Most Attractive Hyderabad Call Girls Kothapet 𖠋 6297143586 𖠋 Will You Mis...
Radiomics Analysis of Pulmonary Nodules in Low Dose CT for Early Detection of Lung Cancer
1. Radiomics analysis of pulmonary
nodules in low-dose CT for early
detection of lung cancer
Wookjin Choi, PhD, Jung Hun Oh, PhD, Sadegh Riyahi, PhD, Feng Jiang, MD, PhD, Wengen Chen, MD, PhD,
Joseph O. Deasy, PhD, and Wei Lu PhD
Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY 10065
Department of Pathology, University of Maryland School of Medicine, Baltimore, MD 21201
Department of Diagnostic Radiology and Nuclear Medicine, University of Maryland School of Medicine, Baltimore, MD 21201
Radiomics and Quantitative Imaging
TH-AB-201-10
2. Lung Cancer Screening
• Early detection of lung cancer by LDCT can reduce
mortality
– LDCT dramatically increases the number of
indeterminate pulmonary nodules (PNs)
• Known features correlated with PN malignancy
– Size, growth rate
– Calcification, enhancement, solidity → texture features
– Boundary margins (spiculation, lobulation) → shape and
appearance features
2
Benign pattern of calcification
Malignant nodules
Benign nodules
Images from radiologyassistant.nl, AJR Am J Roentgenol. 2003 May;180(5):1255-63, and AJR Am J Roentgenol. 2002 May;178(5):1053-7.
3. Data set
A subset of LIDC-IDRI fromTCIA
• Multi-institution data
• Four radiologists detected and contoured
PNs
• Consensus contour: generated by STAPLE
using 2 or more contours of PN
• Biopsy-proven ground-truth or 2 years of
stable PN
• 36 benign and 43 malignant cases, 7 missing
contours (5 benign and 2 malignant)
• 72 cases evaluated (31 benign and 41
malignant cases)
3
LIDC-IDRI: Lung Image Database Consortium image collection, TCIA: The Cancer Imaging Archive,
STAPLE: the simultaneous truth and performance level estimation
Data From LIDC-IDRI. The Cancer Imaging Archive. http://doi.org/10.7937/K9/TCIA.2015.LO9QL9SX
# Pts
Total 1,010
Having diagnosis data 157
Primary cancer
biopsy-proven
progression
43
42
1
Benign
biopsy-proven
2yrs of stable PN
progression
36
7
26
3
Metastatic cancer
or unknown
78
4. ACR Lung-RADS
Category Baseline Screening Malignancy
1 No PNs; PNs with calcification
Negative
<1% chance of malignancy
2
Solid/part-solid: <6 mm
GGN: <20 mm
Benign appearance
<1% chance of malignancy
3
Solid: ≥6 to <8 mm
Part-solid: ≥6 mm with solid component <6 mm
GGN: ≥20 mm
Probably benign
1-2% chance of malignancy
4A
Solid: ≥8 to <15 mm
Part-solid: ≥8 mm with solid component ≥6 and <8 mm
Suspicious
5-15% chance of malignancy
4B
Solid: ≥15 mm
Part-solid: Solid component ≥8 mm
>15% chance of malignancy
4X
Category 3 or 4 PNs with suspicious features (e.g. enlarged lymph nodes)
or suspicious imaging findings (e.g. spiculation)
>15% chance of malignancy
4
Summary of Lung-RADS categorization for baseline screening
ACR: American College of Radiology
Lung-RADS: Lung CT Screening Reporting and Data System
5. Radiomics for Lung Cancer Screening
• Radiomic features from 3D volume and 2D axial slice with largest area
(n=103)
– Shape: 40 features (3D: 26 and 2D: 14)
– Texture: 36 features (GLCM: 16 and GLCM: 20)
– Intensity: 18 features (3D: 9 and 2D: 9)
– Shape+Intensity: 9 features, shape features weighted by intensity using image
moment (3D: 5 and 2D: 4)
5
GLCM: gray level co-occurrence matrix, GLRM: gray level run-length matrix
GLCM GLRM
Texture features Intensity features
3D 2D
Shape features
6. Prediction model
• Distinctive features (n=50)
– Hierarchical clustering using Pearson
correlation
– 9 shape, 26 texture, 8 intensity, and 7
shape+intensity features
– 15 significant features after Bonferroni
correction
• SVM classification coupled with LASSO
feature selection
– Selected 10 most important features by 10-
fold CV of the LASSO
– Radial basis function kernel
(γ = 0.001 and C = 64)
– 10 times 10-fold CV
6
SVM: Support vector machine, LASSO: Least absolute shrinkage and selection operator,
CV: Cross validation
7. Performance of the SVM-LASSO model
7
CV: Cross validation, SVM: Support Vector Machine
with increasing number of features in the 10x10-fold CV
8. using the two important features and compared with Lung-RADS
Performance of the SVM-LASSO models
Prediction Model Sensitivity Specificity Accuracy AUC # of Features
Lung-RADS 73.3% 70.4% 72.2% 0.74 4
SVM-LASSO 10×10-fold 87.9±2.5% 78.2±1.6% 83.7±1.7% 0.86±0.01 2
20×5-fold 86.0±3.3% 75.9±3.9% 81.6±2.6% 0.85±0.02 2
50×2-fold 83.4±4.9% 71.9±8.8% 78.5±5.1% 0.84±0.03 2
8BB: Bounding Box, AP: Anterior-Posterior, SD: Standard Deviation, IDM: Inverse Difference Moment
• BB_AP
– Highly correlated with the axial longest diameter and its
perpendicular diameter (r = 0.96, larger – more malignant)
• SD_IDM
– Directional variation of local homogeneity (smaller – more
malignant)
9. Scatter plot of the two features
9
and the classification curve by the SVM-LASSO model
10. Cases misclassified by Lung-RADS
10
BB: Bounding Box, SD: Standard Deviation, AP: Anterior-Posterior, SI: Superior-Inferior, IDM: Inverse Difference Moment
Scale bar is 10 mm, Spiculation: 1(no)-5(marked) scale
but correctly classified by the SVM-LASSO model
11. Comparison with recent models
Dataset Model description
Hawkins et al.
(2016)
Baseline CT scans of 261pts in
NLST
Biopsy-proven ground-truth or 2
years of stable PN
23 RIDER stable radiomic features
Random forest classifier
10×10-fold CV
Ma et al.
(2016)
LIDC 72pts
Biopsy-proven ground-truth or 2
years of stable PN
583 radiomic features
Random forest classifier
10-fold CV
Buty et al.
(2016)
LIDC 2054 PNs
Ground-truth by radiologist’s
assessment
Spherical Harmonics (100, 150, and 400 shape features)
and AlexNet33 (4096 appearance features)
Random forest classifier
10-fold CV
Kumar et al.
(2015)
LIDC 97pts,
including metastatic tumors
Biopsy-proven ground-truth or 2
years of stable PN
Deep convolutional neural network model (5000
features)
10-fold CV
Proposed
LIDC 72pts
Biopsy-proven ground-truth or
2 years of stable PN
2 important features
LASSO features selection and SVM classification
10×10-fold CV
11
12. Comparison with recent models
Sensitivity Specificity Accuracy AUC
Hawkins et al. (2016) 51.7% 92.9% 80.0% 0.83
Ma et al. (2016) 80.0% 85.5% 82.7%
Buty et al. (2016) 82.4%
Kumar et al. (2015) 79.1% 76.1% 77.5%
Proposed 87.9% 78.2% 83.7% 0.86
12
• A large number of features applied comparing to number of patients
– May cause model overfitting problem
• No discussions on how the selected features might have contributed to
the prediction of malignancy
• Deep learning needs numerous training data to avoid model overfitting,
and transfer learning is questionable
13. Future Works
• Candidate feature approach
– Quantification of spiculated or lobulated margins
– Calcification, attachment, solidity and cavitation of PNs
• Integrate plasma biomakers in the SVM-LASSO model
– Difficult to diagnose small PNs, 50% accuracy when PN size
< 15mm
– Combining plasma biomarkers with clinical variables and image
features (AUC = 0.95)
• Deep learning - Data Science Bowl 2017, Predicting Lung Cancer
– 3D Fully Convolutional Neural Network model
– Ranked 99th out of 1972 teams (Top 6%, Bronze medal)
13
Jiang et al. Int J Cancer. 2017. [published online ahead of print 2017/06/06].
14. Conclusion
• Developed an SVM-LASSO model to predict
malignancy of the indeterminate PNs
– Two important features: the bounding box anterior-
posterior dimension and the standard deviation of local
homogeneity
– The proposed model outperformed Lung-RADS
• A multicenter clinical trial in a large population is
required
– To prospectively and vigorously validate the radiomic
features
– Can be translated into clinical practice
14
18. Results
ROC curve analysis on the best model
of SVM-LASSO and Lung-RADS
The box plots show the difference between benign and malignant
PNs for the selected features (BB_AP and SD_IDM) and the
largest diameter. P-values were obtained by theWilcoxon rank
sum test and adjusted using Bonferroni correction
18
19. BB_AP 10mm
19
Benign
IDM_LR: 0.172
IDM_AP: 0.182
IDM_SI: 0.284
Mean_IDM: 0.174
SD_IDM: 0.033
Malignant
IDM_LR: 0.116
IDM_AP: 0.136
IDM_SI: 0.138
Mean_IDM: 0.111
SD_IDM: 0.014
Axial Sagittal Coronal
d e f
LR
AP
AP
SI
LR
SI
a b c
LR
AP
AP
SI
LR
SI
20. BB_AP 17mm
20
Benign
IDM_LR: 0.276
IDM_AP: 0.316
IDM_SI: 0.210
Mean_IDM: 0.220
SD_IDM: 0.030
Malignant
IDM_LR: 0.234
IDM_AP: 0.215
IDM_SI: 0.236
Mean_IDM: 0.203
SD_IDM: 0.020
Axial Sagittal Coronal
d e f
LR
AP
AP
SI
LR
SI
a b c
LR
AP
AP
SI
LR
SI
Notes de l'éditeur
Lung cancer is the leading cause of cancer death in the world.
The early detection of lung cancer by LDCT can reduce mortality.
Frequent use of LDCT dramatically increases the number of indeterminate pulmonary nodules (PNs) and producing a high false-positive diagnostic rate or overdiagnosis.
Therefore, it is important to develop new approaches to improve the accuracy.
We also performed Lung-RADS categorization based on the PN contour and the physician’s annotations.
The Lung Imaging Reporting and Data System (Lung-RADS) was developed by the American College of Radiology (ACR) to standardize the screening of lung cancer on CT images.
As shown in the Table, the Lung-RADS categorization is mainly based on PN size (the average of the longest and shortest diameters on axial slice)
with some consideration to calcification, appearance type (solid, part-solid, and non-solid or ground glass nodule/GGN), and additional suspicious features.
To match the original LIDC-IDRI diagnosis, categories 3 and lower are labeled as benign and category 4 (4A, 4B, and 4X) as malignant.
79 LDCT scans: 36 benign and 43 malignant cases, 7 missing contours
We performed Lung-RADS categorization based on the PN contour and the physician’s annotations.
Inverse Difference Moment – local homogeneity
SD IDM – directional variance of local homogeneity
least absolute shrinkage and selection operator (LASSO) is a regression analysis method that performs both variable selection and regularization in order to enhance the prediction accuracy and interpretability of the statistical model it produces
Compared with Lung-RADS
This figure shows that the CV achieved the highest accuracy when two features were selected into the SVM classifier.
The two most frequently selected features were: (1) the tumor bounding box (BB) anterior-posterior (AP) dimension (BB_AP), and (2) the standard deviation (SD) of inverse difference moment (SD_IDM).
The best single-feature model achieved 76.0±2.4% accuracy (0.77±0.02 AUC).
When adding the second feature in the model, the prediction accuracy was improved to 83.7±1.7% (0.86±0.01 AUC).
However, the performance got worse when adding two or more features as shown in Figure 2.
Table 2 compares the prediction performance of Lung-RADS and the SVM-LASSO model.
The Lung-RADS achieved an accuracy of 72.2% with four features (size, type, calcification, and suspicious features or image findings).
The SVM-LASSO model achieved an accuracy of 83.7% with two features (BB_AP and SD_IDM), which represented an 11.5% improvement over the Lung-RADS.
The performance of the SVM-LASSO model was stable when more patients were partitioned into the testing set with 5- and 2-fold CVs compared to 10-fold CV.
Only a small reduction in each accuracy measurement was observed, and even the accuracy of 2-fold CV (50% patients were partitioned into the training set and 50% the testing set) was 6% higher than that of the Lung-RADS.
All five malignant PNs in the elliptical region a were misclassified as benign by Lung-RADS since they were small (size<8 mm).
On the other hand, all five benign PNs in the elliptical region b were misclassified as malignant by Lung-RADS since they were large (size≥8 mm).
The SVM-LASSO model correctly classified these ten PNs in both regions a and b by combining a size feature (BB_AP) with a texture feature (SD_IDM).
There was one malignant case that was correctly classified by Lung-RADS but misclassified by the SVM-LASSO model.
This case was indicated by arrow c in Figure 5.
This figure shows five examples of such PNs.
Region a: Two example cases (c) and (d) were shown. Note that case (c) was part-solid with a solid component of 4 mm.
Region b: Two example cases (a) and (b) were shown
This case was indicated by arrow c in Figure 5 and shown as the case (e) in Fig. 6.
This PN was originally classified by Lung-RADS as benign (category 3) based on its size, and then adjusted to be malignant (category 4X) due to its spiculated margins.
Finally, 10 cases were misclassified by both Lung-RADS and the SVM-LASSO model.
Overall, the SVM-LASSO model showed a clear advantage over Lung-RADS.
It should be noted that each method was evaluated on different datasets or with different validation methods.
The proposed method showed comparable or better accuracy than others,
Applying pre-trained model using general images to medical images
a feature discovery approach[1) extracted a large number of image features (>100), 2) selected the most informative ones that are independent, robust, and prominent on the data]
plan to add a candidate feature approach[a few important features are selected based on prior knowledge of their physiological, biochemical or functional associations with the disease and therapy]
a PN with smooth, well-defined margins is more likely benign
while a PN with lobulated or spiculated margins is more likely malignant.
More accurate PN segmentation method is required as well as advanced feature extraction method
We showed that when combining plasma biomarkers with clinical variables and image features, the prediction was very accurate (AUC = 0.95).
These suggested that the biomarkers, clinical variables, and image features had complementary information.
Therefore, we plan to integrate all these information in the SVM-LASSO model and expect further improvement in the prediction accuracy even for small PNs.
Limitations of the present study include that the model was developed from a moderate-size cohort of 72 patients.
Although the CVs showed that the model using the two features was not notably affected by overfitting,
the performance of the model should be validated in a larger, independent patient cohort.
This work was supported in part by the National Cancer Institute Grants R01CA172638.