S 16 - My scientific paper in the field of oncologic imaging

Purpose:

Normal saline infusion radiofrequency ablation (NSRFA) is a widely accepted treatment for tumours up to 3 cm in diameter. Microwave ablation (MWA) and hydrochloric acid infusion radiofrequency ablation (HRFA) may be better alternatives for larger tumours, but few studies comparing these 2 modalities have been published. Thus, this study is aiming to compare ablation zone sizes and shapes resulting from HRFA and MWA, using NSRFA as a control in an ex vivo bovine liver model.

Methods and materials:

A total of 90 ablation procedures were performed prospectively, each using 3 modalities: NSRFA, HRFA, and MWA. For each modality, 5 ablation procedures were performed for each combination of power (80W, 100W, or 120W) and duration (5, 10, 20, 30, 45, or 60 minutes). Ablation zone outcome parameter includes transverse diameter (TD), longitudinal diameter (LD), volume (V), front distance (FD), and spherical ratio (SR).

Results:

For ablation duration up to 30 minutes, the mean TD after HRFA and MWA did not differ significantly (β=0.13, P=.20). For ablation durations greater than 30 minutes, the mean TD was significantly larger after HRFA than after MWA (β=1.657, P<.001). The largest TD (9.46 cm) and the most spherical ablation zone (i.e. highest SR, 0.95) resulted from HRFA performed with 100W power for 60 minutes.

Conclusion:

Both MWA and HRFA with power settings of 80W to 120W and a duration of 30 minutes are capable of treating tumours of 5 cm or less. For tumours of 7 cm or larger, HRFA with power settings 100W and duration of 30 minutes or longer is a better option than MWA.

Limitations:

THe ex vivo model is partly different from the clinical patient.

Ethics committee approval

n/a

Funding:

National Natural Science Foundation of China, No. 81771955.

Purpose:

The pre-operative assessment for lung cancer surgery includes a total body FDG PET/CT as a mandatory investigation for solid and sub-solid lesions.

We sought to evaluate the clinical effectiveness of a PET-CT scan in the management of sub-solid nodules. We additionally assessed the relationship between SUVmax and nodule type.

Methods and materials:

We retrospectively analysed 407 patients who underwent lung resections and received complete imaging work-up. A consultant radiologist confirmed the sub-solid nature of 94 lesions: 27 GGOs and 67 PSN. We then compared the histopathology results with CT and PET-CT.

Results:

Comparing PET/CT staging to pathological staging, PET overstaged (p=0.013) a high proportion of patients with GGOs (42.9%) and PSN (19.4%), whereas the number understaged by PET were not significantly different (p=0.11).

Then, assessing PET/CT and CT concordance, we demonstrated that these investigations were never discordant with GGO and rarely (4.5%) in PSN (p=0.050).

Regression analysis assessed the correlation between SUVmax and nodule type, even when size was taken into account, and both size and nodule type were independent predictors of increasing SUVmax (p<0.001).

Conclusion:

This study suggests that PET/CT rarely adds information for PSN and never for GGO compared with CT, whereas the relationship between SUVmax and nodule type revealed that GGO and PSN had significantly lower SUVmax than solid, suggesting a different behaviour of these lesions.

Thus, there is a strong argument for stopping PET for GGOs and careful consideration in PSN given the low likelihood of change from the CT findings. Further studies should test cost and clinical effectiveness of a PET/CT scan (£89,300 for this cohort of patients) in the management of sub-solid lesions.

Limitations:

The retrospective assessment of data.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To differentiate benign from malignant lung nodules in PET-negative cases based on CT texture analysis (CTTA).

Methods and materials:

We retrospectively analyzed 53 patients with lung nodules with a negative PET scan (standardised uptake value [SUV] <2.5) who underwent a CT-guided lung biopsy (CTLB). Based on the pathological report, the population was divided into two groups: malignant and benign lesions. A dedicated software (TexRAD) was used to segment target nodules, drawing a region of interest (ROI) for each CT slice which displayed the lesion. CT texture parameters, including kurtosis, standard deviation, mean, and skewness were extrapolated for each spatial scale image filtration (SSF) of texture parameters. P<0.05 was statistically significant. The receiver operating characteristic curve (ROC) was calculated considering as dependent variables each of the texture parameters and a dichotomous variable indicating the histological diagnosis (0=malignant, 1=benign) as independent.

Results:

Kurtosis showed significant difference for each SSF (all P≤0.0013; SSF0 1.50±5,70 vs 0.99±4.22; SSF2 1.83±6.18 vs 0.30±3.06; SSF3 1.80±6.81 vs -0.09±1.91; SSF4 1.90± 5.38 vs 0.07±2.65; SSF5 1.3±3.77 vs 0.17±3.33; SSF6 0.88±2.5 vs 0.35±3.67) and was considered the best parameter to differentiate malignant from benign pulmonary nodules. The ROC analysis showed significant area under the curves (AUC) for kurtosis at SSF3 (P<0.001; AUC 0.654), skewness at SSF4 (P<0.001; AUC 0.642), and for standard deviation at SSF0 (P<0.001; AUC 0.821).

Conclusion:

CTTA was useful in identifying significant features that can distinguish malignant from benign pulmonary lesions in cases without an altered glucose metabolism at the PET scan.

Limitations:

The retrospective design.

Ethics committee approval

IRB approved.

Funding:

No funding was received for this work.

Purpose:

Bias field correction deals with the intra-volume intensity inhomogeneities in magnetic resonance imaging (MRI) data. It has little significance for visual diagnosis yet it is a crucial step for the automation of radiotherapy solutions. Extensive research and evaluation show that the often used and well-established correction methods typically require parameter tuning for the given dataset or are expensive to compute.

Methods and materials:

A novel method has been introduced, allowing for the training of a convolutional neural network (CNN) on non-medical images, making the model general and evaluating it for bias field correction on medical MRI data.

Results:

The proposed method has been compared to an optimised N4ITK in three experiments. It produces nearly identical results to N4ITK for a benchmark test and similar results in a coefficient of variation, mean absolute error, mean error, and structural similarity for a BrainWeb-based experiment, with an average speedup factor of 137. Different methods performed best for the different investigated tissue types. It was also tested for real cases, often producing more radio realistic bias fields in the cases where N4ITK struggles. Also, no datasets were encountered where the method would not improve image quality.

Conclusion:

Training a convolutional neural network implicitly, through the relation of the outputs of two similar training samples, is possible, omitting the need for true target data. The achieved model is not only generalised but also achieves comparable accuracy to N4ITK with a significant speedup.

Limitations:

No numerical evaluation for real cases.

Ethics committee approval

n/a

Funding:

Support obtained from the Cancer Research Foundation in Northern Sweden and from Karin and Krister Olsson.

Purpose:

Survival rates and the prognosis of patients with pancreatic adenocarcinoma (PA) directly depends on the histologic grade of the tumour. The purpose of this study was to identify features and quantitative parameters at preoperative CT imaging for predicting the histologic grade of the tumour.

Methods and materials:

81 patients with histologically-verified pancreatic ductal adenocarcinoma (11 PA grade 1, 47 PA grade 2, and 23 PA grade 3) and preoperative CT imaging performed in the native, arterial, and portal phases (10 and 40 seconds since reaching the threshold density value on aorta), with the tube voltage 120 kV, were enrolled in this study. We compared mean absolute tumour and pancreatic parenchyma attenuation between groups for each contrast enhancement phase. For each patient, the presence of rim enhancement was assessed and defined as irregular peripheral hyperdensity with a relatively hypodense central area on dynamic-enhanced images.

Results:

The mean tumour density in the arterial phase was 51.1±17.6 HU, median 49 HU. The mean tumour density in the venous phase was 63.0±18.9 HU, median 63 HU. Among PA grade 1, in 81.8% the rim enhancement was not observed. PA grade 2 was found in 85.71% cases where it was impossible to unequivocally speak out in favour of the absence or presence of the rim enhancement. In the presence of the rim enhancement in 96.42%, PA was grade 2 or grade 3. The density of PA grade 1 in 90.91% was <63 HU and the density of PA grade 3 in 69.57% was ≥63 HU.

Conclusion:

Rim enhancement of PA was associated with more aggressive histologic tumour grades.

Limitations:

n/a

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

In literature, there is little documentation of ADC values in normal bone marrow and of the effects of physiological factors. The aim of this study is to evaluate how sex, age, and fat (evaluated from %FF) affect the ADC values measured in the bone marrow of healthy individuals.

Methods and materials:

We processed the diffusion-weighted images of WB-MRI examinations in 100 asymptomatic individuals, 50 men and 50 women aged 30-79 years. The mean value of ADC in bone marrow (ADCbm) was estimated by fitting the histogram extracted from the images with a semi-automatic segmentation technique. The %FF value was measured as the average of ROIs drawn in specific positions (iliac bone, D10, L4). Gender differences were assessed with the Mann-Whitney U test, while ADCbm correlations with age and %FF were assessed with the Spearman correlation coefficient (ρ).

Results:

The values of ADCbm were significantly higher in women than in men (respectively 458.1±63.3 µm²/s and 383.0±58.4 µm²/s, p<0.01). A significant negative correlation of ADCbm with age (ρ=-0.46, p< 0,001) was observed in women but was absent in men (ρ=-0.04, p=0.79). Moreover, there was a negative correlation between ADCbm and %FF in both women and men (respectively: ρ=-0.52, p<0.001 and ρ=-0.28, p<0.05).

Conclusion:

There was a significant difference between ADCbm values of men and women. The negative correlation between ADCbm and age in women is likely related to menopause, while that between ADCbm and %FF, present in both sexes, appears to be a separate process.

Limitations:

Calculate ADCbm and %FF with two different approaches.

Ethics committee approval

Approved by an internal review board.

Funding:

No funding was received for this work.

Purpose:

HPV-positive and HPV-negative forms of OPSCC are biologically distinct entities, with different prognosis, treatment strategies, and divergent AJCC/UICC staging schemes. Using feature selection and machine-learning classification algorithms, we identified the radiomics signature of HPV on pre-treatment FDG-PET and non-contrast CT scans.

Methods and materials:

All data was retrieved from the Cancer Imaging Archive. Patients with OPSCC, known HPV-status, and pre-treatment PET/CT were included. The primary tumour and metastatic cervical nodes were separately segmented on PET scans; segmentations were then copied and adapted to the coregistered CT. A set of 1,040 radiomics features was extracted from each segmentation and per imaging modality.

To predict HPV-status, minimum-redundancy-maximum-relevance (MRMR) feature selection and random forest (RF) machine learning classifiers were applied in 5-fold cross-validation, repeated x10. The area-under-the-curve of receiver-operating-characteristic-curves (AUC-of-ROC) averaged across validation folds is reported.

Results:

Of 114 included OPSCC cases, 87 had HPV-positive and 27 had HPV-negative forms. In addition to primary tumour lesions, 119 HPV-positive and 51 HPV-negative metastatic cervical nodes were analysed.

The primary tumour RF models achieved an averaged AUC-of-ROC of 0.69 (PET), 0.75 (CT), and 0.73 (PET/CT), based on 20 features selected by MRMR from the respective feature sets. In comparison, the averaged AUC of lymph node RF models was 0.66 (PET) and 0.64 (CT), based on 20 MRMR-features. PET-based lymph node models were non-predictive.

Conclusion:

Radiomics feature extraction from FDG-PET and non-contrast CT scans, combined with machine learning classifiers, can generate imaging biomarkers for HPV-status in OPSCC primary tumours and metastatic lymph nodes, which may aid pathologists in HPV-classification if standard immunohistochemical staining is equivocal or supplement the immunohistochemical tests in subjects requiring second-line testing.

Limitations:

n/a

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

WB-MRI is increasingly recommended for the evaluation of patients with metastatic bone diseases. A semi-automatic technique for the segmentation of diffusion-weighted images in WB-MRI examination has been developed to allow quantitative evaluation of tumour burden in bone metastases. The aim of this study was to evaluate the intra and interobserver repeatability of quantitative analysis of ADC in healthy and metastatic bone marrow in patients with bone metastases from breast cancer (BCa) and prostate cancer (PCa).

Methods and materials:

4 independent observers processed WB-MRI examinations from 20 patients with bone metastases (10 women with BCa and 10 men with PCa). Segmentation of bone marrow was performed applying a threshold on high b-value diffusion-weighted images and manually removing misclassified non-bone regions. The total volume of bone marrow (V_bm) and first-order ADC statistics were then obtained. We measured repeatability using intraclass correlation coefficients (ICC) and Bland-Altman method.

Results:

We observed good to excellent ICC values in 4 parameters: ADC_Mean with intra and interobserver ICCs respectively of 0.92 (0.85-0.96, CI 95%) and 0.85 (0.75-0.92), ADC_Median 0.92 (0.85-0.96) and 0.88 (0.79-0.93), ADC_5% 0.9 (0.81-0.94) and 0.8 (0.69-0.88), and ADC_Skewness 0.9 (0.80-0.94) and 0.86 (0.76-0.92). Bland-Altman bias and limits of agreement calculated for ADC_Mean were, respectively, -1.1% and -13.7%–12.1%, and 5.0% and -18.8%–15.3% in intra and interobserver analysis. Repeatability was higher in BCa patients than in PCa patients.

Conclusion:

While bias was almost 5 times greater for interobserver relative to intra, the limits of agreement were increased by only about 25%.

Limitations:

Difficulties in segmenting PCa patients data likely led to lower repeatability in these patients relative to BCa patients.

Ethics committee approval

Approved by internal review board.

Funding:

No funding was received for this work.