RPS 1405b - Artificial intelligence and CT radiomics

Purpose:

To evaluate the performance of an arterial phase CT radiomics model in predicting tumour progression prior to transplant or post-transplant recurrence in patients with hepatocellular carcinoma (HCC) who underwent pre-transplant transarterial chemoembolisation (TACE)

Methods and materials:

In this retrospective study, we analysed the baseline contrast-enhanced CTs of 91 patients with HCC listed for liver transplantation prior to bridging therapy with TACE. Using the PyRadiomics library (v2.2), 1,441 radiomic features were extracted from the largest HCC volume in the arterial phase. After removing features with variance below 0.05, 762 features remained. Principal component analysis (PCA) was applied for feature reduction and the first 22 components were kept in the model. The binary classification task was done using a support vector machine (SVM) classifier. The performance of the model was evaluated by area under the receiver operating characteristic curve (AUC) using 5-fold cross-validation. The primary endpoint of the radiomics classifier was the combined event of tumour progression prior to transplant or post-transplant recurrence.

Results:

The combined event occurred in 37% of patients (34/91) with 21.9% of patients (20/91) having tumour progression and1 5% of patients (14/91) having tumour recurrence. The mean time from listing to transplant, listing to progression, and listing to recurrence was 182, 335, and 670 days, respectively. The mean follow-up time was 1,308 days. The mean AUC for the prediction of pre-transplant tumour progression or post-transplant tumour recurrence was 0.81 (95% confidence interval: 0.76-0.85). The best radiomic model at a threshold of 0.387 achieved a sensitivity and specificity (at maximum Youden index) of 85% and 83%, respectively.

Conclusion:

An arterial phase-based radiomics model can predict tumour progression and recurrence in HCC-patients listed for transplant who undergo TACE

A machine-learning driven radiomics algorithm may provide a useful tool for risk stratification of HCC patients listed for transplant. Further prospective validation is required.

Limitations:

A retrospective study and single institution cohort. Radiomic features were obtained from single arterial-phase.

Ethics committee approval

An IRB approved study.

Funding:

No funding was received for this work.

Purpose:

To biologically validate the CT-radiomics signature predictive of response to immune-checkpoint-inhibitors by correlating the CT-radiomics score with gene expression profiling data from tumour samples and to compare the performance of radiomics only and combined with clinical data to predict response to immune-checkpoint-inhibitors in patients with advanced solid tumours.

Methods and materials:

In this retrospective study, a predictive signature was derived from 239 metastases of 85 consecutive patients treated with immune checkpoint inhibitors (monoclonal anti-PD-1/anti-PD-L1 antibodies) monotherapy in phase I clinical trials from August 2012 to May 2018. Validation was performed in 112 lesions from 46 consecutive patients with urinary bladder cancer treated with anti-PD-1/PD-L1 monotherapy (Cohort 2). Radiomics variables were extracted from all lesions per patient in the pre-treatment CT scan and an elastic-net model was implemented. Biological validation was pursued studying the association (Mann-Whitney analyses) of the CT-radiomics score with RNA signatures of cytotoxic cells in an independent cohort of 20 patients. A regression model was applied to combine radiomics and clinical variables.

Results:

In Cohort 1 (training), the radiomics signature associates with a response (area under the curve [AUC] 0.74;P=2.96x10^-10). In Cohort 2 (validation), the radiomics signature predicts a response with AUC=0.70 (P=1x10^-3). A high radiomics score associates with cytotoxic-enriched immunophenotype (P=0.035). The model combining radiomics and clinical variables slightly improve the capacity for predicting response AUC=0.76 (P=1.02x10^-5) in the training set and AUC=0.78 (P=5x10^-4) in the validation set.

Conclusion:

CT-radiomics signature at baseline predicts the response to immune checkpoint inhibitors and, likely, reflects tumour immunophenotype. Integrating radiomics and clinical data improve the prediction performance.

Limitations:

The biological validation was performed in a relatively small cohort.

Ethics committee approval

Approved by the institutional review board. Informed consent was waived.

Funding:

Funded by BBVA, La Caixa-Foundation,Prostate Cancer Foundation.

Purpose:

To evaluate repeatability and reproducibility of radiomics features in spectral CT using an anthropomorphic abdomen phantom.

Methods and materials:

The phantom was scanned repetitively using different radiation doses (CTDI_vol of 10, 15, and 20 mGy) using a spectral detector computed tomography scanner (SDCT). Scanning was repeated three times each. Different spectral results were reconstructed: conventional, virtual non-contrast (VNC), iodine density, iodine no-water, MonoE 50KeV, electron density, and Z-Effective. 34 three-dimensional regions-of-interest (ROIs) were defined analytically and fully-automated. 93 radiomics features were extracted employing the pyradiomics-package. In total, 93x7x3x3 extracted features were analysed. Repeatability and reproducibility were assessed using the concordance correlation coefficient (CCC).

Results:

Features with excellent and moderate reproducibility, as indicated by a CCC>0.8 and CCC>0.6, respectively, for the different image reconstructions were identified. Percentage of features with CCC>0.8 varied between 16% (Z-Effective with 10 mGy) and 89% (iodine no-water with 15 mGy). Furthermore, the total number of features with CCC>0.8 and CCC>0.6 increases with dose. In total there were 205 features with excellent repeatability and reproducibility, but only 13 features for all image reconstructions were identified. Among all reconstructions, first-order features show the best reproducibility and repeatability (~50%). CCCs between different radiation doses were further assessed showing good reproducibility (>80% of features).

Conclusion:

This study comprises the first systematic evaluation of repeatability and reproducibility using spectral reconstruction images from SDCT. While first-order features show a good repeatability and reproducibility, care has to be taken when using features with medium and poor reproducibility for classification tasks.

Limitations:

Only phantom images were assessed; a translation onto patients is desirable. Images did not undergo pre-processing which holds potential to improve repeatability and reproducibility.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To develop an immunotherapy response model based on the change of radiomic features from baseline to follow-up for a more accurate response assessment.

Methods and materials:

This retrospective study included 67 consecutive patients with metastatic solid tumours treated in phase I clinical trials with immune checkpoint inhibitors (anti-PD1/PD-L1). Baseline contrast-enhanced CT and follow-up CT after 8 weeks of treatment were evaluated. One target lesion per patient was delineated and radiomics features were extracted using pyradiomics. An analysis of variance (ANOVA) was implemented as a feature selector. Featured differences between baseline and follow-up that showed significant differences for clinical benefit were selected and a logistic regression model was generated. Log-rank analysis of Kaplan-Meier data was conducted to determine associations between the radiomics score and overall survival.

Results:

From 105 radiomic features from first-order, shape, and texture matrices, four features differed between baseline and follow-up (total energy, small area low grey-level emphasis, coarseness, and surface) associated with a response in the validation set (p<0.05). Prediction capacity of the radiomics model showed an AUC of 0.79 (95% CI 0.69-0.90; p<0.05). Patients with a higher radiomics score showed longer survival (p<0.05).

Conclusion:

The change in radiomics features from baseline to follow-up after treatment classifies responder and non-responder patients and associates with overall survival.

Limitations:

A retrospective study with a limited number of patients.

Ethics committee approval

Approved by the ethics committee.

Funding:

No funding was received for this work.

Purpose:

To compare the results of conventional and targeted CT-based radiomic classification models distinguishing benign and malignant pulmonary nodules.

Methods and materials:

The two datasets were comprised of routine and targeted lung CT images of 349 patients with pathological results respectively. We built classification models to predict the malignancy of pulmonary nodules based on a radiomics method. To achieve the best result on both datasets, we used diverse feature selection methods and machine learning algorithms.

Results:

The AUC, accuracy, sensitivity, and specificity of the radiomic model based on routine lung CT images were 0.77, 60.16%, 58.15%, and 78.79%, respectively. By contrast, the AUC, accuracy, sensitivity, and specificity of the radiomic model based on targeted lung CT images were 0.84, 75.72%, 76.36%, and 69.70%, respectively. The result showed that targeted CT images have an edge over routine CT images on predicting malignancy of pulmonary nodules.

Conclusion:

Because of the use of target scanning, we obtain more sufficient and more accurate image data, and the large matrix, small FOV, and thin layer thickness reduce the voxel, making the quantitative analysis results more accurate. We have demonstrated that a radiomic classification model based on targeted CT images has better performance compared to a model based on routine CT images. Therefore, target CT combined with radiomic classification models can provide a more accurate diagnosis of pulmonary nodules, thus helping select proper treatment or follow-up strategies.

Limitations:

This work has not been verified by a multicentre trial, which is considered to be our future work.

Ethics committee approval

This retrospective study was approved by the institutional review board of Shanghai Chest Hospital (KS1956).

Funding:

Interdisciplinary Program of Shanghai Jiaotong University (YG2017QN661), Shanghai Municipal Health Commission Project (20184Y0219)(2019SY063), Science and Technology Commission Shanghai Municipal Project (18511102902)(19411965200), Shanghai Key Laboratory Open Project (STCSM18DZ2270700).