We performed a retrospective analysis of pretreated advanced NSCLC cancer patients treated with Nivolumab between January 2015 and July 2017 at the Radiation Oncology Unit of Siena University Hospital, at the Medical Oncology and Translational Oncology Units of Catanzaro University Hospital and at Reggio Calabria Grand Metropolitan Hospital in Italy.

In our analysis all patients with a complete setting of diagnostic histological samples and a clinical-radiological pre-treatment staging, including a total body CT scan with and without contrast, and an intact (i.e., not previously treated with surgery or radiotherapy) evaluable target lesion in the lung were included.

All the patients gave written consent to anonymous use of their examinations for research scope. A study notification was submitted to local ethical committee as established by national laws. The University Hospital of Siena Institutional Review Board at the pilot center authorized the retrospective analysis of the data. All procedures were undertaken in compliance with the ethical statements of the Helsinki Declaration (2008) of the World Medical Association.

All the patients were staged as IIIB/IV and progressed after systemic therapy with cis-platinum doublets. All the patients received intravenous nivolumab (3 mg/kg in 60 min) on biweekly bases. Treatment was given until progression, unacceptable toxicity or death.

All CTs at Siena University Hospital were performed using a 64-detector row CT scanner (Discovery 750 HD; GE Healthcare, Milwaukee, WI, USA).

In all patients, CT of the chest was performed with a spiral technique in tail-cranial direction (from the bases of the lungs to a plane cutting through the upper thoracic outlet, with the patient lying supine). Enhanced CT scans were obtained in the portal venous phase (delay 65–80 sec) with an intravenous injection of 2 ml/kg of non-ionic contrast material after a bolus injection of non-ionic contrast material (Iopamidolo 370 mgI/ml; Iopamiro, Bracco Diagnostics, Italy or Iopromide 370 mgI/ml, Ultravist 370 mgI/ml; Bayer HealthCare Pharmaceuticals, Italy), followed by 30 ml of saline solution using a peristaltic semiautomated power injector (3–4 ml/sec flow rate; SIAS 757, Bologna, Italy) with an 18-gauge needle in the antecubital vein. The following technical parameters were used: slice thickness 2.5 mm, beam pitch 1.375/0.937, reconstruction interval 0.8 mm, 120–140 kVp and 250–500 mA. An automatic current modulation tube was used to minimize radiation exposure. Standard reconstruction algorithm was used. Patients were instructed not to breathe during helical imaging in order to avoid motion artefacts.

A CT-scan was repeated, in order to assess the response to immunotherapy, every 3 months or in any case showing clinical signs suggesting progressive disease (PD). General examinations with recording of toxicity, blood cell counts and chemistry and serum markers levels were evaluated every two weeks.

The gross tumour volume (GTV) was contoured by a radiation oncologist (VN) and confirmed by three expert radiologists (EA, MAM, VR) on pre-contrast and post-contrast sequences (Fig. 1). The impact of variations on contouring was analysed performing two delineation on each patient, and the TA parameters were evaluated for reliability with Intraclass Coefficient Correlation method (ICC). All the analyses for work have been accomplished with LifeX Software^© (34). TA parameters were limited to features of gray-level co-occurrence matrix (GLCM), sphericity and indices from the gray-level histogram, to avoid an excessive number of parameters and overfitting of models.

The reliability analysis performed with ICC showed that the TA parameters that resulted significantly reproducible (ICC>0.70, single measure) were, respectively, 12 out of 14 for the post-contrast CT (85%), 13 out of 14 (93%) for the pre-contrast CT (Table I).

In order to perform a statistical correlation among the reliable TA parameters and outcome, we used the software X-Tile (35) to calculate for each TA parameter a cut-off value that could be significant at survival analysis (Kaplan Meier analysis), on the overall population, after normalization of the parameters for the training and validation population.

For each significant cut-off, we assigned the score 1 for the subgroup with the worse prognosis and 0 for the subgroup with the better prognosis.

We then summed the scores to obtain a global texture score that was finally partitioned into two subgroups. The clinical variables of the two subgroups (sex, age, histology) were compared with Chi-square test, in order to exclude biases due to clinical factors not taken into account for the comparison of the above-mentioned subsets.

To validate our performance model, our cohort of patients were separated into two partitions, with the patients treated in our Unit (35 patients) as the training dataset and the patients treated in other Institutions (24 patients) as the validation set.

A survival analysis of progression free survival (PFS) and overall survival (OS) with Kaplan-Meier method was used in the two subgroups to test the texture score.

PFS was calculated from the date of the patient's beginning of immunotherapy to date of CT examination showing progression of disease, or censored to last follow-up visit. Conversely, OS was calculated from the date of the patient's beginning of immunotherapy to the death of patients or censored to last follow-up visit.

All the statistical analysis was conducted with SPSS software v.23.0.

This is a retrospective analysis performed on an unmasked sample of fifty-nine consecutive patients with advanced NSCLC cancer who had progressed standard frontline chemotherapy and had received salvage therapy with Nivolumab, between January 2015 and July 2017. We have enrolled 48 (81%) males and 11 (19%) females with a median age of 69 years (mean +-/ SD=67.4 +/- 9.7 years, range 42–86 years). Thirty-six patients (61%) were diagnosed as adenocarcinoma, 18 (30.5%) as squamous cell carcinoma and five as undefined NSCLC histology. Forty-five patients were in stage IV (76%), whereas 14 patients were in stage IIIB (24%). Τhe description of parameters in the two datasets is presented in Table II.

Median follow-up was 12 months and, on the overall, these patients presented a median PFS of 9 months (mean +-/ SD=11 months +/- 1.3 months, 95% CI 8.3–13.7 months), and a median OS of 13 months (mean +-/ SD=14.5 months +/- 1.6 months, 95% CI 11.1–17.4 months), with 32 (54%) patients still alive at the last follow-up examination (Fig. 1).

Median cycles of Nivolumab were 10 for the whole population (mean 14, range 2–52 cycles). In our series, survival analysis did not find any correlation with gender (male: Median OS 13 months, mean 14 +/- 1.8 months, 95% CI 10–18 months, vs. female: Median OS 7 months, mean 10 +/- 2.7 months, 95% CI 5–15 months, P-value: 0.58), with age (Cox-Regression P-value: 0.44), stage (IIIB: Median OS 15 months, mean 14 +/- 2.4 months, 95% CI 10–19 months, vs. IV: Median OS 13 months, mean 12 +/- 1.7 months, 95% CI 9–15 months, P-value: 0.60) and with histology (squamous cell: Median OS 15 months, mean 14 +/- 2.7 months, 95% CI 9–20 months, vs. adenocarcinoma: Median OS 13 months, mean 12 +/- 1.6 months, 95% CI 9–15 months, P-value 0.88).

In order to calculate a reproducible cut-off for the targeted TA parameters in CT images, we used the X-Tile Software as described in the Methods section.

A valid parameter cut-off could be selected in the whole population for: 1) volume (score 1 > 36 ml); 2) histogram entropy (score 1 > 1.30), 3) compacity (score 1 < 3.4), GLCM-entropy (score 1 > 1.8), 5) GLCM-dissimilarity (score 1 > 5.6) GLCM-correlation (score 1<0.5). These cut-off values were applied to both training and validation set.

In our analysis, we have restricted the testing of texture parameters describing morphological, histogram and GLCM features, as these parameters have been successfully used and correlated in lung cancer imaging (26), to avoid the risk of overfitting our model using too many parameters (36).

The global texture score, as described in the Methods section, allowed to classify the patients into two cohorts (low risk: Score 0–1, and high risk: Score >1).

The clinical parameters of the two cohorts were then tested with Chi-square analysis, that demonstrated no statistical differences of the clinical variables in the two subsets, as follows: Age (P=0.282), sex (P=0.843), histology (P=0.107), stage (P=0.834) (Table II).

PFS resulted significant for both the subsets TRAINING SUBSET (low risk median PFS 10 months, mean 10.9 +/- 1.8 months, 95% CI 7–14 months, vs. high risk median PFS 3 months, mean 4.7 +/- 1.1 months, 95% CI 2–6 months, P-value: 0.04) and VALIDATION SUBSET (low risk median PFS 15 months, mean 15.5 +/- 2.9 months, 95% CI 9–21 months, vs. high risk median PFS 6 months, mean 6 +/- 0.9 months, 95% CI 4–8 months, P-value: 0.04; Fig. 2).

OS, also, resulted significant in both subsets TRAINING SUBSET (low risk median OS: Not reached, mean 14.4 +/- 2 months, 95% CI 10–18 months, vs. high risk median OS: 5 months, mean 6.2 +/- 1.6 months, 95% CI 3–9 months, P-value: 0.02) and VALIDATION SUBSET (low risk median OS: 26 months, mean 19 +/- 3 months, 95% CI 13–25 months, vs. high risk median OS 6 months, mean 6.1 +/- 0.9 months, 95% CI 4–8 months, P-value 0.03; Fig. 2).

The global texture score correlated with patients' survival allowed to classify patients included into two cohorts at low (Score 0–1, 36 patients, 61%) and high risk (Score> 1, 23 patients, 39%) of recurrence, respectively, presenting a median PFS of 12 months (mean 13 +/- 1.7 months, 95% CI 9–16 months), vs. a median PFS of 5 months (mean 5.4 +/- 0.7 months, 95% CI 4–7 months), with a P-value of 0.01, and a median OS of 26 months (mean +/- SD: 17.5 +/- 1.98 months, 95% CI 14–21 months) and 5 months (mean +/- SD: 6 +/- 0.99 months, 95% CI: 4–8 months), with a P-value <0.01, in the low and high risk groups, respectively (Fig. 2).

Number of cycles of Nivolumab, also, were different among subgroups (P=0.026), with low risk subset showing a median of 12 cycles (mean 18.5 cycles, range 2–52 cyles) and high risk subset showing a median of 10 cycles (mean 8.8 cycles, range 2–20 cycles).