Copyright
©The Author(s) 2020.
Artif Intell Gastroenterol. Aug 28, 2020; 1(2): 37-50
Published online Aug 28, 2020. doi: 10.35712/aig.v1.i2.37
Published online Aug 28, 2020. doi: 10.35712/aig.v1.i2.37
Table 1 Application of radiomics in qualitative diagnosis in gastroenterology
Classification of disease | Imaging modality | Features evaluated and methods | Outcomes | Ref. |
Gastric disease | ||||
AC; GIST; lymphoma | CECT | RLM; GLCM; absolute gradient; autoregressive model; wavelet transformation | Texture-based lesion classification in arterial phase differentiated between AC and lymphoma, and GIST and lymphoma, with misclassification rates of 3.1% and 0%, respectively Texture-based lesion classification in venous phase differentiated between AC and GIST, and different grades of AC with misclassification rates of 10% and 4.4%, respectively | [30] |
Borrmann type IV GC; PGL | CECT | A total of 485 3D features, divided into four groups: First order statistics, shape and size based features, texture features, and wavelet features | The subjective findings model, radiomics signature, and combined model showed a diagnostic accuracy of 81.43% (AUC, 0.806; sensitivity, 63.33%; specificity, 95.00%), 84.29% (AUC, 0.886; sensitivity, 86.67%; specificity, 82.50%), and 87.14% (AUC, 0.903; sensitivity, 70.00%; specificity, 100%), respectively, in the differentiation of Borrmann type IV GC from PGL | [31] |
Hepatic disease | ||||
Neoplastic and bland portal vein thrombus | CECT | Mean; entropy; SD of pixel intensity; kurtosis; skewness | In the discrimination of neoplastic from bland thrombus, the AUCs were 0.97 for mean value of positive pixels, 0.93 for entropy, 0.99 for the model combining mean value of positive pixels and entropy, 0.91 for thrombus density, and 0.61 for the radiologist's subjective evaluation | [42] |
HCC; MT; HH | MRI | GLCM; GLRLM; GLSZM; NGTDM | Texture analysis in differential diagnosis: HCC and MT: accuracy 92%, sensitivity100%, specificity 84%, AUC 0.95 HCC and HH: accuracy 90%, sensitivity 96%, specificity 84%, AUC 0.95 MT and HH: accuracy 73%, sensitivity74%, specificity72%, AUC 0.75 | [41] |
Pancreatic disease | ||||
PSCN | CECT | A total of 385 radiomics high-throughput features: Intensity; wavelet; NGTDM | The accuracy rate of SCNs before surgery was only 30.4% (31/102) while the diagnostic model established based on dual-phase pancreatic CT imaging features had an improved accuracy rate of diagnosis, showing an AUC of 0.767, sensitivity of 68.6%, and specificity of 70.9% | [51] |
PNEC; PDAC | CECT | Filtration-histogram approach and Laplacian-of-Gaussian band-pass filters (sigma values of 0.5, 1.5, and 2.5) were used and texture parameters under different filters, including: Kurtosis, skewness, entropy, and uniformity | PNEC showed a lower entropy and a higher uniformity compared to PDAC in the portal phase with an acceptable AUC of 0.71-0.72 | [52] |
Colorectal disease | ||||
Neoplastic and non-neoplastic lesions | CECT | 78 features for each lesion in total | Combining high-order CT images with CT volumetric texture features allowed a significantly higher AUC of 0.85 in distinguishing neoplastic colon tumors from non-neoplastic ones than only using the image intensity (AUC of 0.74) | [58] |
Table 2 Application of radiomics in disease staging in gastroenterology
Classification of disease | Imaging modality | Features evaluated and methods | Outcomes | Ref. |
Esophageal disease | ||||
ESCC | Unenhanced CT and CECT | Six parameters based on HU values: Mean; 10th percentiles; 90th percentiles; kurtosis; entropy; skew | Kurtosis and entropy based on unenhanced CT were an independent predictor of T stages, lymph node metastasis (N- vs N+), and overall stages Skew and kurtosis based on unenhanced CT images showed significant differences among N stages as well as 90th percentile based on contrast-enhanced CT images; entropy and 90th percentile based on CECT images showed significant correlations with N stage and overall stage | [18] |
ESCC | CECT | A total of 9790 radiomics features were extracted including the following four categories: First-order histogram statistics, size and shape-based features, texture features, and wavelet features | The radiomics signature significantly associated with ESCC staging and yielded a better performance for discrimination of early and advanced stage ESCC compared to tumor volume | [19] |
Gastric disease | ||||
GC | MRI | Entropy-related parameters based on ADC maps including: (1) First-order entropy; (2–5) second-order entropies, including entropy(H)0, entropy(H)45, entropy(H)90, and entropy(H)135; (6) entropy(H)mean; and (7) entropy(H)range | All the entropy-related parameters showed significant differences in gastric cancers at different T, N, and overall stages, as well as at different statuses of vascular invasion Entropy, entropy(H)0, entropy(H)45, and entropy(H)90, showed significant differences between gastric cancers with and without perineural invasion | [35] |
GC | CECT | Mean; maximum frequency; mode; skewness; kurtosis; entropy | Maximum frequency in the arterial phase and mean, maximum frequency, mode in the venous phase correlated positively with T, N, and overall stage of GC; entropy in the venous phase also correlated positively with N and overall stage; skewness in the arterial phase had the highest AUC of 0.822 in identifying early from advanced GCs | [34] |
Hepatic disease | ||||
Hepatic fibrosis | CECT | Mean gray-level intensity; entropy; kurtosis; skewness | Mean gray-level intensity, mean, and entropy increased with fibrosis stage; kurtosis and skewness decreased with increasing fibrosis | [43] |
Pancreatic disease | ||||
PNET | CECT | Positive pixels; SD; kurtosis; skewness; entropy | Entropy was predictive of Grades 2 and 3 tumors with an accuracy rate for classifying G1, G2, and G3 tumors of 79.3% | [53] |
PNET | CECT | Mean value; variance; skewness; kurtosis; entropy | Kurtosis was significantly different among the three G groups, giving an AUC value of 0.924 for the diagnosis of G3 with a sensitivity and specificity of 82% and 85%, respectively; entropy differed significantly between G1 and G3 and between G2 and G3 tumors, giving an AUC value of 0.732 for the diagnosis of G3 with a sensitivity and specificity of 82% and 64%, respectively | [54] |
Colorectal disease | ||||
CRC | CECT | The 16-feature-based radiomics signature was generated using LASSO logistic regression model | The 16-feature-based radiomics signature was an independent predictor for staging of CRC and could categorize CRC into stage I-II and stage III-IV Compared with the clinical model, the radiomics signature showed significantly better performance either in the training dataset (AUC: 0.792 vs 0.632; P < 0.001) or in the validation dataset (AUC: 0.708 vs 0.592; P = 0.037) | [59] |
Table 3 Application of radiomics in evaluation of therapeutic efficacy and prognosis in gastroenterology
Classification of disease | Imaging modality | Features evaluated and methods | Outcomes | Ref. |
Esophageal disease | ||||
EC | 18F-FDG PET | A total of 38 features (such as entropy, size, and magnitude of local and global heterogeneous and homogeneous tumor regions) extracted from 5 different textures | Tumor textural analysis provided non-responder, partial-responder, and complete-responder patient identification with a higher sensitivity (76%-92%) than any SUV measurement | [22] |
ESCC | MRI | 138 radiomic features were extracted from each image sequence based on three principal methods: Histogram-based (IH, GH), texture-based (GLCM, GLRLM, and NIDM), and transform-based (GWTF) | Radiomic analysis showed that CR vs SD, PR vs SD, and responders (CR and PR) vs non- responders could be differentiated by 26, 17, and 33 features, respectively; the prediction models (ANN and SVM) based on features extracted from SPAIR T2W sequence (SVM: 0.929; ANN: 0.883) showed higher accuracy than those derived from T2W (SVM: 0.893; ANN: 0.861) | [24] |
Gastric disease | ||||
GC | CECT | Histogram features: Kurtosis, skewness; GLCM: ASM, contrast, entropy, variance, correlation | Contrast, variance, and correlation showed fair accuracy for the prediction of good survival with all AUCs being over 0.7, and all were statistically significant | [38] |
Hepatic disease | ||||
HCC | CECT | 21 textural parameters per filter were extracted from the region of interests delineated around tumor outline by application of a Gabor filter and wavelet transform with 3 band-width responses (filter 0, 1.0, and 1.5) | Texture analysis was observed to have potential in assessment of prognosis and selection of appropriate patients with intermediate-advanced HCC treated by TACE and sorafenib | [46] |
HCC | CECT | First order statistics; geometry; texture analysis; GLCM | Textures derived from pretreatment dynamic CT imaging were analyzed, higher arterial enhancement ratio and GLCM moments, smaller tumor size, and lower tumor homogeneity were significant predictors of CR after TACE | [47] |
Pancreatic disease | ||||
Pancreas head cancer | CECT | Laplacian of the Gaussian band-pass filter was applied to detect intensity changes within the images smoothened by Gaussian distribution based on the filter sigma value of 1.0 (fine texture, filter width 4 pixels), 1.5 to 2.0 (medium texture, filter width 6-10 pixels), and 2.5 (coarse texture, filter width 12 pixels) | Texture parameters of average, contrast, correlation, and standard deviation with no filter, and fine to medium filter values, as well as the presence of nodal metastasis were significantly different between recurred and non-recurred patients; lower standard deviation and contrast and higher correlation with lower average value representing homogenous texture were significantly associated with poorer DFS, along with the presence of lymph node metastasis | [55] |
PDAC | CECT | Mean gray-level; intensity; entropy; MPP; kurtosis; SD; skewness | Tumor size, tumor SD, and skewness were significantly and independently associated with PFS, while tumor size and tumor SD were significantly and independently associated with OS | [56] |
Colorectal disease | ||||
LARC | MRI | 18 features extracted using the Haralick's GLCM and 12 parameters calculated for the histogram-based analysis | Radiomics based on pre-treatment and early follow-up MRI could provide quantitative information to differentiate pCR from non-pCR, and GR from non-GR. | [60] |
Rectal cancer | MRI | Kurtosis; entropy; skewness; MPP | The change in kurtosis between midtreatment and pretreatment images was significantly lower in the PR + NR subgroup compared with the pCR subgroup; pretreatment AUROC to discriminate between pCR and PR + NR, was significantly higher for kurtosis (0.907, P < 0.001) | [61] |
- Citation: Feng P, Wang ZD, Fan W, Liu H, Pan JJ. Diagnostic advances of artificial intelligence and radiomics in gastroenterology. Artif Intell Gastroenterol 2020; 1(2): 37-50
- URL: https://www.wjgnet.com/2644-3236/full/v1/i2/37.htm
- DOI: https://dx.doi.org/10.35712/aig.v1.i2.37