Radiomics: a Study of Outcome in Lung Cancer

1. Treatment of lung cancer Lung cancer is the most common cause of cancer death in The Netherlands, with an annual incidence exceeding 8,000. Two main variants of lung cancer can be identified: small cell and non-small cell lung cancer (SCLC and NSCLC respectively), the latter comprising approximately 80% of the lung cancer cases. Despite treatment improvement, the prognosis of NSCLC remains poor, with a median survival of 8 months after diagnosis, and a 5 year survival of less than 13%. Radiotherapy plays a key role in the treatment of NSCLC. Over the years, radiotherapeutical treatment options have increased tremendously. These include dose escalation, more intensive schedules and concurrent chemo-radiotherapy. These schedules have improved both local control and survival in patients. However, they also induce more toxicity, and the radiation oncologist faces the challenging task of choosing the optimal therapy for each patient: taking into account tumor characteristics as well as the patient's condition. In other words: the physician must estimate the expected therapeutic ratio often on a background of insufficient outcomes information. The same problem arises in other therapies for lung cancer, chemotherapy and surgery. In early disease, surgery is the mainstay of the treatment of lung cancer patients. This can be combined with neoadjuvant or adjuvant chemotherapy and/or radiotherapy.
2. Prediction of response A major problem in lung cancer management is the lack of data dealing with predictive factors for prognosis and treatment outcome. The currently used staging system (TNM) does not accurately predict outcome within homogeneous treatment groups. As a result, an individualized therapeutic ratio cannot be calculated, leading to either over- or under-treatment of many patients and hampering further optimization of any therapy. Attempts were undertaken to refine and improve the risk stratification, leading to the development of several prediction models. The performance of the models is usually expressed as the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC). The maximum value of the AUC is 1.0; indicating a perfect prediction model. A value of 0.5 indicates that patients are correctly classified in 50% of the cases, e.g. as good as chance. While high prediction accuracy (AUC=0.85) has been achieved for a population of NSCLC patients of all stages, treated with different modalities, it is a more challenging task to predict survival accurately when focusing on a subgroup. A pretreatment prediction model for patients treated with surgery yielded an AUC of 0.61 while a pretreatment model for patients treated with radiotherapy resulted in an AUC of 0.75. Further improvement of the radiotherapy model was obtained by adding information about blood biomarkers and this extended model yielded an AUC of 0.83. Investigating other blood biomarkers or possible combinations of biomarkers is a challenge and our results underline the importance of using these data in addition to clinical and imaging parameters. Survival remains certainly an outcome of major importance, but the last decades other treatment-related outcome measures, such as radiation induced lung injury or esophageal damage, became more important for the evaluation of treatment results. Pneumonitis or radiation induced lung injury has been subject of many studies. However, results are quite difficult to interpret, because many different variables, dosimetric parameters as well as other treatment or patient related characteristics, have been identified, studies showed inconsistent or even conflicting results, and sample sizes were often very limited. Recently, Chen et al. published a neural network model for prediction of grade 2 or higher pneumonitis, which yielded an AUC of 0.74 in the test dataset31. Compared to other models, these results are promising, although external validation of the model is warranted before it can be used in clinical practice. Our group developed a model predicting dyspnea ≥ grade 2 according to CTCv3.0. Patient as well as dosimetric parameters were incorporated in the model, which resulted in a cross-validated AUC of 0.62. In summary, existing models perform rather well, but there is a lot of room for improvement by adding new factors as well as applying advanced model building techniques. Prediction models still have to be developed for a number of clinically relevant outcomes. Finally, incorporating confidence intervals in the prediction as well as quantifying the gain in prediction precision if a certain diagnostic/ prognostic test is performed, would certainly be of great value for clinical use of the models (http://www.predictcancer.org/).
3. Strategies to improve prediction models for lung cancer

In order to improve the prediction models for survival as well as toxicity outcome one can include many variables as possible predictors including imaging, genomics and proteomics information.

3.1 Imaging

An important feature for prognosis on the FDG-PET-scan is the maximal Standardized Uptake Value (SUVmax). There is a statistically significant difference in 2-year survival between patients with a high pretreatment SUV and a low pretreatment SUV. Patients with a low SUVmax had a 2-year survival of 90.6%, while patients with a high SUVmax had a 2-year survival of only 58.6%. There is a significant correlation between high SUVmax and a high HIF1α staining in the biopsies, which is a marker for hypoxia. Non significant relations were shown for CA IX, Ki67 and Glut-1 and SUVmax.

Besides FDG, new PET-tracers are being developed. One of the new tracers is HX4, which is a hypoxia tracer. Regulation of tissue oxygen homeostasis is critical for cell function, proliferation and survival. Evidence for this continues to accumulate along with our understanding of the complex oxygen-sensing pathways present within cells. The microenvironment of tumors in particular is very oxygen heterogeneous, with hypoxic areas, which may explain much of our difficulty in treating cancer effectively. This is true when comparing levels of hypoxia among different patient tumors, but also within individual tumors. Accumulating evidence implicates the biological responses to hypoxia and the alterations in these pathways in cancer as important contributors to overall malignancy and treatment efficacy. This has recently prompted several investigations into the possibility of imaging and targeting treatment at the biological responses to hypoxia.

3.2 Gene signatures

Analysis of gene signatures can help to improve the predictive value of the model. An example of this, is the proliferation signature investigated by Starmans et al. Two different signatures of 110 genes were compared in prognostic value. Both showed a very good prognostic value on breast cancer data sets. The AUC (area under the curve) improved when the proliferation signature were added to the models of clinical factors. Another gene profile was tested on early stage NSCLC. This profile consists of 72 genes and is validated on stage I and II NSCLC patients of five centers. It was possible to identify early-stage NSCLC patients with high and low risk for disease recurrence and death within 3 years after primary surgical treatment.

3.3 Tumor biopsies

Hypoxia is (besides in serum) also measurable in the tissue itself. Several markers of hypoxia are predictive for survival. An example is HIF1α, which is upregulated is case of hypoxia. A higher staining of HIF1α is correlated with a worse prognosis in NSCLC. CA IX correlated with severe and chronic hypoxia, and has a strong association with a poor outcome in NSCLC.

Another marker is Ki67, which is expressed in proliferating cells. A higher Ki67 indicates more proliferation, and in a systemic review of Martin et. al. a worse prognosis was shown when Ki67 expression is increased.

3.4 Application of machine learning techniques

The availability of genomic data, together with improved imaging modalities, leads to unprecedented amounts of biological and medical data, which can only be dealt with using computational methods, not only for storing the data, but also for integrating, analyzing, displaying and eventually understanding it.

Machine learning offers a number of techniques for these purposes. These techniques can overcome problems encountered with conventional statistical methods especially if data is highly correlated, many variables are available but a limited number of patients (high-dimensional data), or many different models have to be tested for their predictive value. In the field of radiotherapy and especially for the prediction of treatment responses, machine learning is an upcoming modality. Successes over traditional statistics have already been published 43and first promising results for building predictive models concerning survival of non-small-cell lung-cancer are already found in the literature.