Historically, medical imaging has been a qualitative or semi-quantitative modality. It is difficult to quantify what can be seen in an image, and to turn it into valuable predictive outcomes. As a result of advances in both computational hardware and machine learning algorithms, computers are making great strides in obtaining quantitative information from imaging and correlating it with outcomes. Radiomics, in its two forms “handcrafted and deep,” is an emerging field that translates medical images into quantitative data to yield biological information and enable radiologic phenotypic profiling for diagnosis, theragnosis, decision support, and monitoring. Handcrafted radiomics is a multistage process in which features based on shape, pixel intensities, and texture are extracted from radiographs. Within this review, we describe the steps: starting with quantitative imaging data, how it can be extracted, how to correlate it with clinical and biological outcomes, resulting in models that can be used to make predictions, such as survival, or for detection and classification used in diagnostics. The application of deep learning, the second arm of radiomics, and its place in the radiomics workflow is discussed, along with its advantages and disadvantages. To better illustrate the technologies being used, we provide real-world clinical applications of radiomics in oncology, showcasing research on the applications of radiomics, as well as covering its limitations and its future direction.

In this review we describe recent developments in the field of radiomics along with current relevant literature linking it to tumor biology. We furthermore explore the methodologic quality of these studies with our in-house radiomics quality scoring (RQS) tool. Finally, we offer our vision on necessary future steps for the development of stable radiomic features and their links to tumor biology. Two authors (S.S. and H.W.) independently performed a thorough systematic literature search and outcome extraction to identify relevant studies published in MEDLINE/PubMed (National Center for Biotechnology Information, NCBI), EMBASE (Ovid) and Web of Science (WoS). Two authors (S.S, H.W) separately and two authors (J.v.T and E.d.J) concordantly scored the articles for their methodology and analyses according to the previously published radiomics quality score (RQS). In summary, a total of 655 records were identified till 25-09-2017 based on the previously specified search terms, from which n = 236 in MEDLINE/PubMed, n = 215 in EMBASE and n = 204 from Web of Science. After determining full article availability and reading the available articles, a total of n = 41 studies were included in the systematic review. The RQS scoring resulted in some discrepancies between the reviewers, e.g. reviewer H.W scored 4 studies ≥50%, reviewer S.S scored 3 studies ≥50% while reviewers J.v.T and E.d.J scored 1 study ≥50%. Up to nine studies were given a quality score of 0%. The majority of studies were scored below 50%. In this study, we performed a systematic literature search linking radiomics to tumor biology. All but two studies (n = 39) revealed that radiomic features derived from ultrasound, CT, PET and/or MR are significantly associated with one or several specific tumor biologic substrates, from somatic mutation status to tumor histopathologic grading and metabolism. Considerable inter-observer differences were found with regard to RQS scoring, while important questions were raised concerning the interpretability of the outcome of such scores.

In this study, we investigate the role of radiomics for prediction of overall survival (OS), locoregional recurrence (LRR) and distant metastases (DM) in stage III and IV HNSCC patients treated by chemoradiotherapy. We hypothesize that radiomic analysis of (peri-)tumoral tissue may detect invasion of surrounding tissues indicating a higher chance of locoregional recurrence and distant metastasis. Two comprehensive data sources were used: the Dutch Cancer Society Database (Alp 7072, DESIGN) and “Big Data To Decide” (BD2Decide). The gross tumor volumes (GTV) were delineated on contrast-enhanced CT. Radiomic features were extracted using the RadiomiX Discovery Toolbox (OncoRadiomics, Liege, Belgium). Clinical patient features such as age, gender, performance status etc. were collected. Two machine learning methods were chosen for their ability to handle censored data: Cox proportional hazards regression and random survival forest (RSF). Multivariable clinical and radiomic Cox/ RSF models were generated based on significance in univariable cox regression/ RSF analyses on the held out data in the training dataset. Features were selected according to a decreasing hazard ratio for Cox and relative importance for RSF. A total of 444 patients with radiotherapy planning CT-scans were included in this study: 301 head and neck squamous cell carcinoma (HNSCC) patients in the training cohort (DESIGN) and 143 patients in the validation cohort (BD2DECIDE). We found that the highest performing model was a clinical model that was able to predict distant metastasis in oropharyngeal cancer cases with an external validation C-index of 0.74 and 0.65 with the RSF and Cox models respectively. Peritumoral radiomics based prediction models performed poorly in the external validation, with C-index values ranging from 0.32 to 0.61 utilizing both feature selection and model generation methods. Our results suggest that radiomic features from the peritumoral regions are not useful for the prediction of time to OS, LR and DM.

Hypoxia—a common feature of the majority of solid tumors—is a negative prognostic factor, as it is associated with invasion, metastasis and therapy resistance. To date, a variety of methods are available for the assessment of tumor hypoxia, including the use of positron emission tomography (PET). A plethora of hypoxia PET tracers, each with its own strengths and limitations, has been developed and successfully validated, thereby providing useful prognostic or predictive information. The current review focusses on [18F]-HX4, a promising next-generation hypoxia PET tracer. After a brief history of its development, we discuss and compare its characteristics with other hypoxia PET tracers and provide an update on its progression into the clinic. Lastly, we address the potential applications of assessing tumor hypoxia using [18F]-HX4, with a focus on improving patient-tailored therapies.

Tumor hypoxia increases resistance to radiotherapy and systemic therapy. Our aim was to develop and validate a disease-agnostic and disease-specific CT (+FDG-PET) based radiomics hypoxia classification signature. A total of 808 patients with imaging data were included: N=100 training/ N=183 external validation cases for a disease-agnostic CT hypoxia classification signature, N=76 training/ N=39 validation cases for the H&N CT signature and N=62 training /N=36 validation cases for the Lung CT signature. The primary gross tumor volumes (GTV) were manually defined by experts on CT. In order to dichotomize between hypoxic/well-oxygenated tumors a threshold of 20% was used for the [18F]-HX4-derived hypoxic fractions (HF). A random forest (RF)-based machine-learning classifier/ regressor was trained to classify patients as hypoxia-positive/ negative based on radiomic features. A 11 feature "disease-agnostic CT model" reached AUC's of respectively 0.78 (95% confidence interval [CI], 0.62 to 0.94), 0.82 (95% CI, 0.67 to 0.96) and 0.78 (95% CI, 0.67 to 0.89) in three external validation datasets. A "disease-agnostic FDG-PET model" reached an AUC of 0.73 (0.95% CI, 0.49 to 0.97) in validation by combining 5 features. The highest "lung-specific CT model" reached an AUC of 0.80 (0.95% CI, 0.65 to 0.95) in validation with 4 CT features, while the "H&N-specific CT model" reached an AUC of 0.84 (0.95% CI, 0.64 to 1.00) in validation with 15 CT features. A significant survival split (P=0.037) was found between CT-classified hypoxia strata in an external H&N cohort (n=517), while 117 significant hypoxia gene-CT signature feature associations were found in an external lung cohort (n=80). The disease-specific radiomics signatures perform better than the disease agnostic ones. By identifying hypoxic patients our signatures have the potential to enrich interventional hypoxia-targeting trials.

Hypoxic head and neck tumors respond poorly to radiotherapy and can be identified using gene expression profiles. However, it is unknown whether treatment outcome is driven by acute or chronic hypoxia. Gene expression data of 398 head and neck cancers was collected. Four clinical hypoxia profiles were compared to in vitro acute and chronic hypoxia profiles. Chronic and acute hypoxia profiles were tested for their association to outcome using Cox proportional hazard analyses. In an initial set of 224 patients, scores of the four clinical hypoxia profiles correlated with each other and with chronic hypoxia. However, the acute hypoxia profile showed a stronger association with local recurrence after chemoradiotherapy (p = 0.02; HR = 3.1) than the four clinical (chronic hypoxia) profiles (p = 0.2; HR = 0.9). An independent set of 174 patients confirmed that acute hypoxia is a stronger prognostic factor than chronic hypoxia for overall survival, progression-free survival, local and locoregional control. Multivariable analyses accounting for known prognostic factors substantiate this finding (p = 0.045; p = 0.042; p = 0.018 and p = 0.003, respectively). In conclusion, the four clinical hypoxia profiles are related to chronic hypoxia and not acute hypoxia. The acute hypoxia profile shows a stronger association with patient outcome and should be incorporated into existing prediction models.