RPS 616b - Metastatic malignancies: advanced diagnosis and radiomics data

Purpose:

To validate a machine learning algorithm to predict the response of individual liver metastases (LM) in heavily pretreated HER2 amplified metastatic colorectal cancer (mCRC) patients undergoing dual-target therapy within a phase II clinical study.

Methods and materials:

38 patients with amplified HER2 mCRC enrolled in the multicentre HERACLES trial (NCT03225937) were randomly divided into a training set (28 patients) and a validation set (10 patients). All received anti-HER2 treatment (trastuzumab plus lapatinib) and underwent CT examination every 8 weeks until disease progression. The largest diameter of each LM was measured before and after 3 months of treatment to classify metastases in responders (R+) and non-responders (R-). All LM with a diameter >10 mm were manually contoured on the baseline and 24 texture features were extracted. After selecting the most discriminative features using a genetic algorithm, a Gaussian Naïve Bayes classifier was trained on the training set and applied to the validation dataset.

Results:

The dataset was composed of 141 metastases: 108 (75R+, 33R-) in the training set and 33 (21R+, 12R-) in the validation set. The classifier reached a sensitivity, specificity, PPV, and NPV of 89%, 85%, 93%, and 78% for the training set and 90%, 42%, 73%, and 71% for the validation set, respectively. Difference variance and homogeneity were the two most important radiomics features in our model.

Conclusion:

Our study demonstrated the feasibility of developing a radiomics model able to predict the likelihood of response of individual LM in patients with HER+ mCRC undergoing dual-target therapy.

Limitations:

The low number of lesions in the validation dataset, low specificity, especially in the validation dataset, and the manual segmentation with consequent interobserver variability.

Ethics committee approval

Ethics committee approval obtained.

Funding:

AIRC 5x1000 2018-ID.21091, AIRCx1000-ID 9970, and contribution from F Hoffmann-La Roche.

Purpose:

Radiomics is actually at the forefront of precision medicine, but to fully develop it requires multicentre collaboration. In this context, the impact of inter-reader contouring variability on radiomics features (RFs) must be evaluated.

Methods and materials:

The liver metastases (n=70) from colorectal cancers were segmented on CT scans by two readers. RFs from grey-level co-occurrence and run-length matrices were extracted from the whole-lesion volume (3D ROIs) and from the largest cross-section (2D ROIs) for each metastasis. Inter-reader contouring variability was evaluated with the Dice coefficient (DC) and Hausdorff distance (HD), while its impact on RFs was assessed using the mean relative discrepancy and intraclass correlation coefficient (ICC). For the main lesion of each patient, one reader also delineated circular ROIs in the same slice of the 2D ROIs.

Results:

According to the similarity indices, the best inter-reader contouring agreement was observed for 2D ROIs (mean DC: 0.83±0.09; mean HD: 0.25±0.18). Comparing RFs values, the mean relative discrepancy was lower from 2D than 3D ROIs for 24/32 (75%) RFs, ranging from 0.04%-1,567% and from 0.02%-752%, respectively. Correspondingly, an ICC >0.9 was observed for more RFs from 2D ROIs (53%) than from 3D ones (34%). The variability between 2D and circular ROIs was equal or lower than inter-reader variability for 94% of RFs.

Conclusion:

As automatic segmentation tools are currently lacking, a careful selection of the contouring methods is required to minimise differences in the results of texture analysis. 2D contouring seems to reduce this variability but it could be less informative than a whole-lesion analysis.

Limitations:

The choice of different preprocessing parameters could impact on the results.

Ethics committee approval

All patients signed informed consent.

Funding:

Supported by a grant from AIRC 5 x mille.

Purpose:

Quantification of tumour heterogeneity potentially captures anti-vascular treatment effects beyond mean attenuation and size change during anti-angiogenic drug therapy. We aimed to assess global and locoregional heterogeneity changes during tyrosine kinase inhibitor therapy in metastatic renal cell carcinoma patients within the multicentre STAR trial.

Methods and materials:

59 prospective adults (54 male; mean age 67.5±7.0) with 108 lesions underwent arterial-phase contrast-enhanced CT prior to (T0) and 3 months after treatment (T1). Volumetric regions-of-interest were delineated for the same lesions (5 maximum) at each time point. 183 radiomic features (including surface, shape, fractal, histogram, and second and high-order features) were derived using validated in-house software. Highly correlated features (r>0.80) were excluded prior to further analysis. Changes on treatment were compared using the Wilcoxon signed-rank test (correcting for multiple comparisons, p<0.0025).

Results:

Accompanying a tumour volume decrease on treatment (100.22 cm³ vs 74.77 cm³, p<0.001), global tumour attenuation also decreased (41.6 vs 38.2, p<0.001) while negative skewness improved (-1.4 vs -0.94, p<0.001) in keeping with lower vascularisation. Additionally, locoregional GLCM contrast (37.89 vs 48.77, p<0.001), GLSZM intensity variability (0.003 vs 0.004, p<0.001), and GLTDM texture strength (2.52 vs 3.78, p<0.001) increased. GLRL long-run high-grey-level emphasis (1,857 vs 1,611, p<0.001) decreased, indicating greater local texture variation and higher texture visibility and granularity.

Conclusion:

Following tyrosine kinase inhibitor therapy, a global reduction in tumour attenuation and negative skewness during the arterial phase is accompanied by increasing locoregional variability capturing the spatial heterogeneity of therapy anti-vascular effects.

Limitations:

The small sample and heterogeneity in CT parameters.

Ethics committee approval

Ethical board approval and informed consent - NRES Committee REC ref: 11/NW/0246.

Funding:

The National Institute for Health Research Health Technology Assessment Programme (project number 09/91/21).

Purpose:

To assess the effectiveness of PET-CT with 18-fluorodeoxyglucose (FDG) in patients with carcinoma of an unknown primary origin (CUP).

Methods and materials:

187 patients diagnosed with CUP were included in the single-centre retrospective study from September 2018-March 2019: m/f ratio was 123/64 (65.8%/34.2%) and the mean age of patients was 61.9±7.5 years. All patients had a biopsy of at least one metastatic lesion and the malignant nature of the neoplasm was histologically verified. The standard oncological evaluation (including body CT) done before the PET/CT failed to detect the primary tumour.

Results:

Among 187 patients included in the study, the biopsy-proven histological diagnoses were squamous cell cancer (n=87, 46.5%), melanoma (n=15, 8%), undifferentiated carcinoma (n=45, 24.1%), adenocarcinoma (n=23, 12.3%), and undifferentiated malignant neoplasm (n=17, 9.1%). After PET-CT, the primary tumour was detected in 93 (49.7%) patients. New metastatic lesions were revealed in 93 (49.7%) patients. Re-classification of the TNM stage after PET/CT was recorded in 131 (70.1%) cases. TNM stage reclassification was significantly more often observed in patients with an identified primary tumour (100%) compared to the group with an unknown location of the primary tumour (40.4%, p<0.01).

Conclusion:

PET-CT with 18F-FDG allows the performance of significantly more accurate diagnostic workup and tumour staging in the majority of patients with CUP. PET-CT with 18F-FDG is an essential part of the diagnostic protocol in patients with CUP.

Limitations:

A single-centre study.

Ethics committee approval

The study was approved by the local ethics committee.

Funding:

No funding was received for this work.