Skip to main content
. 2022 Oct 31;12(11):2644. doi: 10.3390/diagnostics12112644

Table 3.

Characteristics of different studies on the use of AI in outcome prediction and response to treatment, including strengths and limitations.

Authors Country Imaging Modality Patient Number Study Nature AI System Validation Main Theme Strengths Limitations
Aerts et al. [104] USA CT 47 Prospective Prognostic radiomics signature Yes Radiomic data could define a response phenotype for NSCLC patients treated with Gefitinib therapy Strong associations - Limited sample size
- Only 11 independent radiomic features
Gevaert et al. [87] USA CT 186 Predictive radiogenomics decision model Association between ground glass opacity and the presence of EGFR mutations Need for validations
Zhao et al. [88] China CT 637 Predictive radiomics model Yes Radiomics-based nomogram, incorporating clinical characteristics, CT features and radiomic features, can non-invasively and efficiently predict the EGFR mutation status Sample size - Different CT scanning parameters
- Single center study
Yamamoto et al. [89] USA CT 172 Retrospective Predictive radiogenomics model Yes ALK+ tumors have a CT radiophenotype that distinguishes them from tumors with other NSCLC molecular phenotypes Multi-institutional, international study cohort - Limited sample
- No treatment response validation
Song et al. [90] China CT 937 Retrospective Three blocks deep learning neural network DLM trained by both CT images and clinicopathological information could effectively predict the ALK fusion status and treatment response - Small size of the ALK-target therapy cohort (n = 91)
Chang et al. [92] China PET/CT 526 Prospective Three predictive radiomics models PET/CT-clinical model has a significant advantage to predict the ALK mutation status - Images acquired and processed in the same way
- Single medical center
Wei et al. [93] China CT 134 Prospective Predictive radiomics signature model via binary logistic regression model The radiomics model (21 features) was superior to clinical model in predicting the efficacy of chemotherapy in patients with SCLC
Bourbonne et al. [96] France 167 Retrospective Three predictive radiomics models via neural network training In patients with lung cancer treated with RT, radiomic features extracted from 3D dose maps seem to surpass usual models based on clinical factors and DVHs in predicting APT and LPT
Jiang et al. [99] China PET/CT 399 Predictive radiomics models via logistic regression and random forest classifiers Yes, five-fold cross-validation Imaging-derived signatures could classify expression rate of specific PD-L1 type - Stage IV NSCLC patients composed a very small proportion
- PET/CT data were obtained in clinical routine through two different manufacture-derived machines with different scanning parameters
Yoon et al. [100] South Korea CT 153 Retrospective Two predictive radiomics model via multivariate logistic regression Quantitative CT radiomic features can help predict PD-L1 expression - Patients were identified only from those having PD-L1 testing results
- Proposed prediction model did not undergo external validation
- PD-L1 test lacks universal reference standards
Khorrami et al. [112] USA CT 139 Prospective Machine learning-based radiomics texture features (DelRADx) Yes DelRADx features were (1) predictive of response to ICI therapy, (2) prognostic of improved overall survival, and (3) associated with TIL density on corresponding diagnostic biopsy samples - Validation in two independent test sets
- Radiomic features extracted also from the annular perinodular regions
- The sizes of cohorts, both for discovery and validation, were relatively small
- Radiomic feature expressions might be sensitive to lesion annotation accuracy
Trebeschi et al. [102] Netherlands CT 203 Machine learning-based radiomics model Yes Higher levels of surface-area-to-volume ratio in nonresponding lesions in both cancers suggest that more compact and spherical profiles are associated with better response Individual lesion-based approach, avoiding the issue of mixed response Need for validation in larger cohorts
Grove et al. [105] USA CT 109 Retrospective Predictive CT-based features: convexity morphological feature and
Quantitative imaging biomarkers can be used as an additional diagnostic tool in management of lung adenocarcinomas. Development of imaging features that were descriptive and reproducible using retrospectively acquired clinical scans - Cohort sample sizes
- The two cohorts are likely not comparable (different overall survival trend)
Tang et al. [106] USA CT 190 Retrospective Immunopathology-informed model (IPIM) Yes First radiomics model to leverage immunopathology features (CD3+ cell density and percent tumor cell PDL1 expression) to obtain immune-informed radiomics model yielded subtypes associated with OS - Conducted at a single institution
Wang et al. [107] USA/China CT 18232 Prospective Fully automated artificial intelligence system (FAIS) Yes FAIS learned to identify patients with an EGFR mutation who are at high risk of having TKI resistance - Other genes are relevant to targeted therapy (e.g., ALK, KRAS)
- Combined method (whole lung + tumor-based) wasn’t studied
Jiao et al. [108] USA CT 421 Convolutional AE DL model with three layers of CNNs Integrating DL radiomics models and CTC counts improves patient stratification in predicting recurrence outcomes for patients treated with SBRT for ES-NSCLC