2.6. El Documento Individual de Adaptación Curricular (DIAC)
2.6.1. Evaluación Psicopedagógica Integral en el Marco de la Educación
The potential of radiomics as a non-invasive biomarker in head and neck cancer has been recognized [84]. As part of this PhD project, I have trained and validated a CT radiomics based prognostic model for local tumor control in head and neck squamous cell carcinoma (HNSCC) (Chapter 3.5). I have observed that tumors more heterogeneous in CT density are at higher risk of recurrence. These findings are consistent with a previously published CT model for overall survival by Aerts et al. [69]. Additionally, the prognostic power of my model (concordance index CI = 0.78) was higher than the prognostic power of CT radiomics overall survival model (CI = 0.69). This supports the hypothesis presented in Chapter 2 that primary tumor heterogeneity is not sufficient to describe such a complex endpoint as overall survival [122]. In my training cohort, the prediction of local tumor control only slightly improved when CT radiomics was combined with clinical prognostic factors (tumor stage, volume and HPV status). In the validation cohort CT radiomics showed higher prognostic power than a combined model. Additionally, the CT radiomics model was also prognostic in the subgroup of HPV negative tumors, with a significant split into two risk groups.
In the next study, I have compared the prognostic power of CT and 18F-FDG PET radiomics in the same cohort of HNSCC (Chapter 3.6). I have found, that the round tumors (based on PET signal autosegmentation) with a focused region of high FDG uptake surrounded by a rim of low FDG uptake were associated with better prognosis. In my early study (Chapter 3.1), I have observed that some of the PET measures depended on tumor stage, for example SUV mean. In contrary, my PET radiomics model was not tumor stage dependent. In the comparison to CT radiomics, the radiomics analysis of tumor metabolism using 18F-FDG PET did not improve the prediction of local tumor control. Both models showed similar prognostic power in the validation cohort. Additionally, the combination of CT and PET radiomics did not improve the modeling. Nevertheless, the PET radiomics model showed a better calibration in the validation set in comparison to the CT model. Some of the misclassified patients were affected by CT artifacts in the tumor region, which might have affected the calibration. Further, the analysis of 3 months post-treatment 18F-FDG uptake also did not improve the prediction of local tumor control (Chapter 3.3) in comparison to pre-treatment imaging. In that study the radiomics was calculated in the location of primary tumor (the primary tumor contours were transferred to post-treatment imaging using rigid registration). A deformable registration or an independent primary tumor bed definition on post-treatment imaging could be considered in the future to improve the predictions.
In the context of correlation between tumor biology and radiological phenotype, to the best of my knowledge, my study investigating HPV CT radiomics signature was the first one in HNSCC with an independent validation (Chapter 3.5). My results obtained on two independent cohorts of patients, supported the hypothesis from exploratory studies [99, 100] that HPV-positive tumors are more homogenous in CT density.
The CT radiomics signature for HPV status was independent from my CT radiomic signature for local tumor control, despite the fact that they both point to tumor heterogeneity. The local control signature identifies more heterogeneous tumors as a group with worse prognosis. In the same line, the HPV
Chapter 4 Discussion and outlook: Radiomics as a non-invasive imaging biomarker
signature indicates that more heterogeneous tumors are HPV negative and this subgroup is known to be linked to worse prognosis. However, these signatures comprised different radiomic features and the predictions obtained from both models were not correlated. For example, the CT radiomics signature for HPV status prediction was valid in the subcohorts of patients with and without local tumor control (Chapter 3.5 supplement).
In the future, our group plans to further validate the obtained results in external datasets. Especially, the PET radiomics results may be biased by single institution data considering the influence of scanner calibration on the conversion to activity measurements [123]. We also plan to investigate in more detail, the unexpected results suggesting that PET radiomics does not improve the predictions of local tumor control in comparison to CT radiomics. Our results were obtained based on the analysis of autosegmented PET volume. It is well-known that parts of tumors do not exhibit increased metabolism. Possibly, PET radiomics performed on the entire GTV will improve local tumor control predictions. Metal artifacts are a common concern in head and neck cancer imaging. In my CT radiomics projects, I have excluded the CT slices affected by artifacts from the analysis. However, a worse calibration in CT radiomics model vs PET radiomics model correlated with the presence of artifacts in some on the misclassified cases. Possibly, my artifacts exclusion technique biased the CT radiomics results. Future studies investigating the impact of the artifacts on the modeling (percentage of volume affected by artifacts or iterative reconstruction algorithms for artifacts reduction) are needed.
In the context of endpoint selection, I have shown that my CT radiomics local tumor control model is more prognostic than overall survival (OS) model [69] and loco-regional control (LRC) model [104]. Recently, Leger et al have also reported that CT radiomics has a higher prognostic power for prediction of LRC than OS [104]. This observation holds true also for the PET radiomics. I have found a significant association between PET radiomics and local tumor control, whereas Vallières et al have not observed any association with LRC [105]. In the future, we plan to investigate whether the prediction of more complex endpoints can be improved by inclusion of lymph node radiomics to LRC modeling and other non-tumor related factors (overall patient status, alcohol consumption, tobacco abuse and age) to OS modeling.
In conclusion, the results from this PhD project support the importance of radiomics in outcome modeling in HNSCC and indicate the areas for further research in this field. Finally, we plan to continue our efforts to find a link between the radiological tumor phenotype with tumor biology via the correlation of radiomic features with histopathology (cell proliferation, microvessl density or hypoxia).
Chapter 4 Discussion and outlook: Heterogeneity of head and neck cancer and its implication on radiotherapy
138