Review of the Role of Radiomics in Tumour Risk Classification and Prognosis of Cancer

Breast Cancer

Breast cancer subtypes are assigned based on tumour histopathologic and molecular information. Despite earlier detection and diagnosis with more advanced imaging modalities, failure of radical treatment and cancer recurrence have contributed to significant mortality in breast cancer patients (3). Like PCa, tumour heterogeneity poses a challenge to prognosis prediction and suitability of treatment.

With complex gene expression in breast tumours, heterogeneity has been studied and analysed for its association with prognosis and risk of recurrence. As multiple driver mutations in breast cancer are dynamic and alter with time, there is an increasing need to assess heterogeneity for better prognosis prediction and treatment guidance (4). Current gene expression profiles are progressively going beyond features observed at conventional histopathologic examination to provide more information on tumour biology and distinguish between breast cancer tumour types, and ultimately improve prediction of recurrence and relevant clinical outcomes (4).

Prostate Cancer

Poor risk stratification can greatly impede clinical outcomes in PCa. Overtreatment of indolent PCa and undertreatment of aggressive tumours are not uncommon. There is little evidence to differentiate patients proceeding with surgery or adjuvant RT against merely surveillance (5, 6). Reports have also shown that overtreatment in a large percentage of clinically indolent patients has led to significant toxicities (7). Conversely, undertreatment of aggressive radioresistant tumours could result in poor local control and higher morbidity risk (5). A US study has reported an associated 1.7-fold risk increase of non-organ confined disease after radical prostatectomy in patients with more than 12 months deferral of treatment (8). Similarly, despite radical treatment, up to 40% of patients with clinically curable intermediate-risk disease will recur due to undertreatment (5). This highlights that current clinical models for risk classification could be improved.

Radiomics

Radiomics is the process of converting digital medical images into mineable high-dimensional data by extracting high-throughput quantitative features (9). Exponential advancement in the field of medical image analysis has accelerated the growth and development of radiomics (10), and radiomics has been shown to be a promising tool in providing comprehensive characterisation of tumour biology (11, 12). When combined with statistical tools, biological models can be developed to potentially improve current prediction methods and accuracy of predicted clinical outcomes, far more than conventional analytical models (12).

Radiogenomics is concerned with the relationship between radiomic features extracted and the underlying molecular features at genomic level, which may improve identification of the underlying biological basis of imaging phenotypes (12). Its goal in RT is to improve stratification outcomes and provide better risk assessment, thereby allowing enhanced radiation therapy care (11). The two major approaches include correlating imaging features with specific genotype or molecular phenotype of tumours, and imaging phenotype with biological underpinnings (12). Genomic analyses can increase our understanding of the heterogeneity of tumours and potentially improve prediction of clinical outcomes.

Despite the potential for greater biological understanding of tumours with the introduction of radiomics, the use of radiomics-extracted quantitative data to provide insight on mechanisms at genetic and molecular levels is still debatable due to the limited evidence available and variables affecting the sensitivity of radiomics tools (13). The aim of this review is to identify if radiomics can improve the prediction of tumour risk classification and prognosis for prostate and breast cancer.

Search Strategy for Identification of Studies

Using the preferred reporting items for systematic reviews and meta-analyses statement (PRISMA), a comprehensive radiomics literature search was conducted on PubMed, EMBASE, MEDLINE and Cochrane databases. Publication dates were limited to 10 years.

The following keywords were used in the search strategy: (“radiomics” OR “radiogenomics”) AND (“prostate cancer OR “breast cancer”) AND (“prognosis” OR “survival” OR “predict*” OR “accuracy” OR “tumour response” OR “tumour biology” OR “tumour type” OR “tumour characterisation” OR “classification” OR “stratification”).

Full text studies were reviewed to identify studies fulfilling the predefined criteria. Eligible studies and their reference lists were screened and reviewed for other potential studies in the field. Review, comparative studies, clinical trials (both randomised controlled and non-randomised) of all phases were screened. The full texts of these studies were then reviewed to identify studies fulfilling the predefined inclusion and exclusion criteria.

Studies containing overlapping or insufficient data for extraction were excluded during screening. Reference lists and related studies/articles in each identified publication were also screened and reviewed to avoid missing relevant studies. Duplicates in the search results were removed.

Outcome Measures

The outcome measures from the selected studies included predictability of the level of tumour risk classification, recurrence risk and disease-free survival (DFS).

Statistical Analysis

Studies used receiver-operating characteristic curve analysis to determine the cut-off point of radiomics signature (Rad-score) for risk classification.

Univariate and multivariate Cox proportional hazards model and Kaplan-Meier curves were used to determine the association of radiomics imaging features with prognosis. Associations between radiomics and predictability of cancer prognosis were evaluated by reviewing the area under ROC curves (AUC).

Quality Analysis

Quality analysis of selected papers was performed using Downs and Black's checklist (14) (Table I). This checklist was used to assess the quality of the studies and to elucidate evidence from quantitative studies for quality assessment. Studies were assessed based on their power calculation performance using 5 domains: 1) study quality, 2) external validity, 3) study bias, 4) confounding and selection bias and 5) power of study. The studies were evaluated and assigned a score (out of 28) corresponding to their level of quality: excellent [26-28]; good [20-25]; fair [15-19]; and poor [≤14].

The literature search yielded 63 relevant publications for inclusion, of which 19 were identified as suitable for further evaluation (Figure 1). No randomised control trials were found. The studies included were published between the years of 2010 and 2019. The sample sizes ranged from 49 to 381 participants.

Prediction of Tumour Risk Classification

In assessing the accuracy of radiomics features in predicting tumour risk classification, 3 studies (15-17) showed promising results. All 3 studies extracted radiomic features from MRI images.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

Flow of search methodology.

Li et al. (16) and Monti et al. (17) demonstrated the correlation between image-based tumour phenotypes and molecular classification of breast tumours. Significant AUC values and high discriminative power in radiomic models were reported. Regarding the accuracy of radiomics in classifying prostate tumours and predicting Gleason score (GS), Chaddad et al. (15) reported the best performing AUC values with combined joint intensity matrix (JIM) and the grey level co-occurrence matrix (GLCM) radiomic features to predict for GS.

Comparing the performance of radiomic models to conventional classification methods, 3 studies (13, 18, 19) illustrated that radiomic models resulted in improved classification results by comparing the performance of MRI radiomics to Prostate Imaging Reporting and Data System (PI-RADS). All reported significant AUC values and an improvement in the overall accuracy of prediction with radiomic models. Chen et al. (18) reported a significantly higher AUC for differentiating high-grade from low-grade PCa for the radiomics-based model as compared to that of PI-RADS v2. Additionally, Wang et al. (19) added that a combination of MR radiomics and PI-RADS yielded higher AUC values than PI-RADS alone.

One of the major discussion points with respect to radiomics research is whether current tumour risk classification methods already yields satisfactory results, thereby making the requirement for radiomics research in this area redundant. In assessing the accuracy of radiomics in predicting tumour risk classification in this review, 3 studies (20-22) highlighted the satisfactory results in predicting risk classification achieved by existing available pathologic classifiers. Studies used Breast Imaging Reporting and Data System (BI-RADS) and PI-RADS as risk classifiers. Bellolio et al. (20) reported the presence of cancer on mammography in different BI-RADS categories and the positive and negative predictive values of BI-RADS classification, which were 55% and 92% respectively. Hamoen et al. (21) and Zhang et al. (22) both conducted a meta-analysis of PI-RADS for PCa and both reported high sensitivity and specificity of PCa detection with PI-RADS.

Crivelli (23) also reported the inadequacies present in current radiomic feature extraction and data sharing and the challenges faced by radiologists in this area.

Comparing the performance of radiomic models to conventional classification methods, Bonekamp et al. (24) argued that radiomics is not superior as the calculated AUC for the mean apparent diffusion coefficient (ADC) showed no significant difference in lesion characterisation performance when compared to radiomic machine learning.

Predictors of Breast and Prostate Cancer Prognosis

In assessing the use of radiomics to predict the prognosis of breast and prostate patients, 5 studies (4, 25-28) showed that promising results achieved by radiomic features. These studies focussed on the use of radiomics to predict the biochemical recurrence of disease and disease-free survival (DFS).

Three studies (4, 25-26) investigated the association between computer-extracted breast MRI phenotypes and the prognosis of breast cancer. Huang et al. (25) employed a combination of PET and MRI radiomics and reported a significant mean AUC in distinguishing recurrence-free survival. Similarly, Li et al. (4) also reported significant AUC in differentiating between good and poor prognosis. Sutton et al. (26) observed a statistically significant correlation of the overall model with Oncotype Dx RS and a significant Spearman's rank correlation coefficient, suggesting that image-based features are promising in predicting disease recurrence. Similarly, Park et al. (27) demonstrated the potential of radiomics in predicting DFS in invasive breast cancer patients. The radiomics nomogram achieved a higher C-index than the clinicopathological or Rad-score-only nomograms.

The prostate cancer study by Gnep et al. (28) on the association of radiomics and biochemical recurrence following RT also agreed that radiomics has potential in predicting disease prognosis. Haralick textural features were reported to have significant correlation with Gleason score and biochemical recurrence.

However, we must be cognisant that the need for radiomics to improve prediction of prognosis is challenged by existing prognostic tools that are already available. Wishart et al. (29) reported that the current prognostication model PREDICT for early breast cancer was discriminative and well-validated. The differences in overall actual and predicted mortality were low and not statistically significant. Significant AUC value for the model was observed. In an updated version of PREDICT, Candido et al. (30) demonstrated a further improved prognostication and treatment benefit model. PREDICT v1 and v2 were reported to have similar AUC for estrogen receptor (ER)-negative disease, but v2 had slightly higher AUC value than v1 for ER-positive disease.

In prostate cancer, Verma et al. (31) also challenged the need for radiomics as a prognostic tool by reporting that prostate specific antigen (PSA) density is itself a strong predictor and significant sensitivity and specificity can be achieved by these models.

Prediction of Tumour Risk Classification

Overall, the results demonstrate that there is a growing body of literature supporting the potential of radiomics in improving risk classification for breast and prostate cancer (Table II).

The accuracy of radiomics models in predicting risk classification was supported by 3 studies (15-17). Li et al. (16) and Monti et al. (17) showed that image-based extracted phenotypes and radiomics features were promising in discriminating breast cancer subtypes and histological outcomes, respectively.

Li et al. (16) also found statistically significant associations between tumour phenotype and the respective receptor status. Aggressive tumours were observed to be larger, more irregular and more heterogeneous in contrast enhancement. Radiomics features, therefore, can characterise imaging phenotypes such as heterogeneity and contrast enhancement, which provide insights into tumour pathophysiologic characteristics. This could thus improve quantitative imaging assessment of tumours, broaden the potential for more accurate prognosis prediction and facilitate more personalised treatment management.

The use of radiomics to characterise intratumoural heterogeneity is supported by Monti et al. (17). Skewness and entropy were observed to be the most recurrent features in the radiomics predictive models employed in that study. This indicates randomness in tumour pathophysiological characteristics and highlights the importance of analysing tumour heterogeneity to differentiate between cancer subtypes.

Chaddad et al. (15) utilised joint intensity matrix (JIM) to translate image heterogeneity into texture predictors that were found to be associated with Gleason score (GS). The difference variance feature extracted from JIM was identified to have the greatest predictive power of GS, as variation between textures was clearly encoded. Such a radiomics approach shows promise in allowing greater understanding of the relationship between intensity values in multi-parametric images. Radiomics can account for cellular heterogeneity within the confirmed biopsy to give a more accurate GS score.

Comparing the performance of radiomic models to conventional classification methods, most studies supported that radiomics models have greater predictive value and could be combined with existing predictive models in some cases to further enhance that predictive power.

When comparing the performance of MR radiomics to PI-RADS in the case of tumour risk stratification, 3 studies (13, 18-19) reported better performance in radiomics models. PI-RADS standardises clinical reporting for consistent interpretation of multiparametric MRI (mp-MRI) in prostate cancer, however, inter-reader variation among radiologists limits the reproducibility of the results and the clinical applicability of PI-RADS as a reliable tool (32). In fact, Schimmöller et al. (33) reported a PI-RADS score of only moderate to good for inter-reader agreement when blinded. Therefore, there is a consensus among the studies that radiomics models outperformed PI-RADS because they can include tumour characteristics, which are imperceptible to the eye, hence, giving a higher efficacy in predictive performance.

Additionally, Wang et al. (19) reported that a combination of MR radiomics and PI-RADS improved the predictive performance of PI-RADS. This indicates an increased potential of radiomics in diagnosing better and stratifying clinically relevant PCa, therefore enabling clinicians to provide personalised clinical treatments for patients.

However, the findings of Algohary et al. (13) are limited in generalisability as the PI-RADS score of 3 cases were excluded from the study group intentionally. These cases had lesion characteristics that were not well-described and can give rise to potentially significant inter-observer variability (34). Therefore, the exclusion of the PI-RADS score of the 3 cases disregarded the major clinical challenge associated and was a limitation in the study.

For BI-RADS, Wanaporn and Ornsiri (35) reported wide variability in the interpretation of breast imaging. Radiologist experience and prior knowledge of BI-RADS guidelines accounted for the variability in agreement.

The accuracy of radiomic models in predicting risk classification is challenged by three studies (20-22), which argue that current methods of classifying tumours based on PI-RADS and BI-RADS have satisfactory sensitivity and specificity. Bellolio et al. (20) studied the predictive value of BI-RADS, which is employed to standardise breast image reporting, and drew a high positive predictive value for BI-RADS classification 4 and 5. However, on deeper analysis, this correlation from BI-RADS could be specific to the respective centre and cannot be generalised, since protocols, radiologist assessments and techniques utilised for biopsies differ. A similar protocol could be performed at various centres to validate this correlation of BI-RADS with tumour classification.

Both meta-analyses reported high sensitivity and specificity for the use of PI-RADS and demonstrated that PI-RADS is promising in accurately detecting PCa (21, 22). However, both studies also reported that significant heterogeneity is present in the calculation of the overall PI-RADS score. Hameon et al. (21) observed that studies with low concerns regarding the applicability of PI-RADS presented higher sensitivity and specificity, whereas those with high concerns presented lower sensitivity and specificity. This contrast proposes that more accurate use of PI-RADS could result in the improvement of the overall accuracy for PCa detection. Hence, there is value attached to PI-RADS and it can potentially predict PCa with high accuracy.

Considering radiologists' reviews on radiomics, Crivelli et al. (23) reported that the use of radiomics could be limited by the lack of an existing standardised system for radiomic feature extraction and data sharing. The lack of understanding of basic radiomics concepts among radiologists could also hinder the routine application of radiomics in the clinical setting. This coincides with the previous 2 meta-analyses which emphasised the importance of training radiologists in using radiomics and evaluate their respective learning curves (21, 22).

Comparing the performance of radiomic models to conventional classification methods, Bonekamp et al. (24) found that although radiomic machine learning performed better than radiologist assessment, it was only comparable and did not outperform mean ADC assessment. No added benefit of radiomics was observed compared to ADC. Hence, radiomics cannot be said to be superior in predicting tumour classification.

Prediction of Cancer Prognosis

The potential of radiomic models in predicting biochemical recurrence of disease as well as disease-free survival is supported by 5 studies (4, 25-28). Huang et al. (25) and Li et al. (4) established the potential of radiomics features in the prediction of prognosis with statistically significant AUC reported in their studies. Huang et al. (25) demonstrated a significant relationship between PET and MRI radiomics clusters and tumour grade. In fact, the potential prognostic value of breast cancer tumour grade for predicting disease survival rate was also reported and supported by Rakha et al. (36). This study focused on the histological grade of tumour and resulted in improved breast cancer classification and staging.

Huang et al. (25) also observed that PET and MRI radiomic features together have greater predictive potential then MRI radiomics alone. This could be valuable in deciphering breast cancer phenotypes and imaging biomarkers show promise in the prediction of disease prognosis. Similarly, in evaluating the risk of breast cancer recurrence, Li et al. (4) reported good differentiation between good and poor breast prognosis using quantitative MRI radiomics. Various gene-assay models were studied, and MRI phenotypes were selected from multiple linear regression analyses. The advances in gene expression profiling have brought about a greater understanding of the complexity within breast tumours and are useful in relating breast cancer expression profiles to prognosis and risk of disease recurrence (37).

The potential of radiomic models in predicting biochemical recurrence of disease is also reflected through the combination of imaging phenotypes with genomic data. Sutton et al. (26) correlated imaging phenotype with genomic information to improve the understanding of genetic variability and thus the ability to predict breast cancer prognosis across the different subtypes.

In predicting individualised DFS estimation in breast patients, Park et al. (27) reported that a combined radiomics-clinicopathological nomogram better predicted DFS outcome than the clinicopathological or Rad-score-only nomograms. This demonstrated the potential of radiomic nomograms in predicting individualised DFS estimation in breast patients. The development and validation of radiomics signature-based nomograms in the preoperative prediction of lymph node metastasis in colorectal cancer (38) and prediction of DFS in early-stage non-small cell lung cancer (39) have already been completed with promising study results. Furthermore, Park et al. (27) also demonstrated that a combined radiomics-clinicopathological nomogram gives better prognostic performance with a higher C-index and superior calibration. This is consistent with findings by Liu et al. (9) which suggest that better prognostic performance is attained when clinicopathological characteristics and radiomic features are used together to predict for sentinel lymph node metastasis. This further enhances the predictive power of radiomic nomograms in estimating disease prognosis.

In prostate cancer, Gnep et al. (28) demonstrated the potential of radiomics in predicting biochemical recurrence. Strong association of Haralick features with biochemical recurrence following prostate RT was observed, which is promising in aiding clinical managements when intensifying or de-intensifying treatments to achieve optimal care. In fact, the use of Haralick features for prediction of disease prognosis and progression in glioblastoma was conducted by Yang et al. (40), where Haralick features were found to be predictive of molecular subtypes and survival status in glioblastoma. This supports the feasibility of using tumour-derived imaging features to predict disease prognosis.

Similar to the use of radiomics in the classification of tumours, the actual need for a tool such as radiomics to improve prognosis prediction can be considered challenged by existing prognostic tools available for breast and prostate cancers. Wishart et al. (29) demonstrated the prediction capability of the current prognostication model PREDICT for early breast cancer. The PREDICT model is validated and highly discriminative and Candido et al. (30) reported encouraging results of the updated model along with improved overall calibration and discrimination.

In prostate cancer, Verma et al. (31) have discussed current prognostic tools. PSA density was reported to have strong discriminative power and is a potential predictor of different indices of aggressive prostate cancer. This could potentially reduce unnecessary biopsies and improve patient's care flow.

Limitations and Challenges of Radiomics

There are several challenges and limitations of radiomics that must be addressed to establish its practicality in routine implementation. First, radiomics feature quantification is highly sensitive to acquisition modes and feature extraction methods. Variation across these models and methods can influence feature quantification and radiomic outcomes (41).

Second, inter-reader variability in some radiomic studies can influence accuracy of results. A number of feature extraction algorithms are user-dependent, and this could affect the reproducibility and stability of results obtained. A study by Saha et al. (42) showed that inter-reader variability in radiomics features has also contributed to the instability of these features and questioned the capability in improving tumour classification. It was reported that the average inter-reader stability for all radiomics features was 0.8474 (95%CI=0.8068-0.8858).

Third, as many of the current studies on radiomics involve a small cohort size, this may introduce bias to study results as a result of higher variability. More radiomic studies with larger sample sizes are needed to clearly define whether there is a clinical benefit in using radiomics.

Finally, the association between the imaged characteristics of tumours and actual tumour biology is indirect and complex. Although many relevant radiomic studies show promising results and statistical correlations between radiomic features and genetic phenotypes, the association cannot be directly inferred as causation.