The reproducibility and predictivity of radiomic features extracted from dynamic contrast-enhanced computed tomography of hepatocellular carcinoma.

PURPOSE: To assess the reproducibility of radiomic features (RFs) extracted from dynamic contrast-enhanced computed tomography (DCE-CT) scans of patients diagnosed with hepatocellular carcinoma (HCC) with regards to inter-observer variability and acquisition timing after contrast injection. The predictive ability of reproducible RFs for differentiating between the degrees of HCC differentiation is also investigated. METHODS: We analyzed a set of DCE-CT scans of 39 patients diagnosed with HCC. Two radiologists independently segmented the scans, and RFs were extracted from each sequence of the DCE-CT scans. The same lesion was segmented across the DCE-CT sequences of each patient's scan. From each lesion, 127 commonly used RFs were extracted. The reproducibility of RFs was assessed with regard to (i) inter-observer variability, by evaluating the reproducibility of RFs between the two radiologists; and (ii) timing of acquisition following contrast injection (inter- and intra-imaging phase). The reproducibility of RFs was assessed using the concordance correlation coefficient (CCC), with a cut-off value of 0.90. Reproducible RFs were used for building XGBoost classification models for the differentiation of HCC differentiation. RESULTS: Inter-observer analyses across the different contrast-enhancement phases showed that the number of reproducible RFs was 29 (22.8%), 52 (40.9%), and 36 (28.3%) for the non-contrast enhanced, late arterial, and portal venous phases, respectively. Intra- and inter-sequence analyses revealed that the number of reproducible RFs ranged between 1 (0.8%) and 47 (37%), inversely related with time interval between the sequences. XGBoost algorithms built using reproducible RFs in each phase were found to be high predictive ability of the degree of HCC tumor differentiation. CONCLUSIONS: The reproducibility of many RFs was significantly impacted by inter-observer variability, and a larger number of RFs were impacted by the difference in the time of acquisition after contrast injection. Our findings highlight the need for quality assessment to ensure that scans are analyzed in the same physiologic imaging phase in quantitative imaging studies, or that phase-wide reproducible RFs are selected. Overall, the study emphasizes the importance of reproducibility and quality control when using RFs as biomarkers for clinical applications.

A quantitative analysis of the improvement provided by comprehensive annotation on CT lesion detection using deep learning.

BACKGROUND: Data collected from hospitals are usually partially annotated by radiologists due to time constraints. Developing and evaluating deep learning models on these data may result in over or under estimation PURPOSE: We aimed to quantitatively investigate how the percentage of annotated lesions in CT images will influence the performance of universal lesion detection (ULD) algorithms. METHODS: We trained a multi-view feature pyramid network with position-aware attention (MVP-Net) to perform ULD. Three versions of the DeepLesion dataset were created for training MVP-Net. Original DeepLesion Dataset (OriginalDL) is the publicly available, widely studied DeepLesion dataset that includes 32 735 lesions in 4427 patients which were partially labeled during routine clinical practice. Enriched DeepLesion Dataset (EnrichedDL) is an enhanced dataset that features fully labeled at one or more time points for 4145 patients with 34 317 lesions. UnionDL is the union of the OriginalDL and EnrichedDL with 54 510 labeled lesions in 4427 patients. Each dataset was used separately to train MVP-Net, resulting in the following models: OriginalCNN (replicating the original result), EnrichedCNN (testing the effect of increased annotation), and UnionCNN (featuring the greatest number of annotations). RESULTS: Although the reported mean sensitivity of OriginalCNN was 84.3% using the OriginalDL testing set, the performance fell sharply when tested on the EnrichedDL testing set, yielding mean sensitivities of 56.1%, 66.0%, and 67.8% for OriginalCNN, EnrichedCNN, and UnionCNN, respectively. We also found that increasing the percentage of annotated lesions in the training set increased sensitivity, but the margin of increase in performance gradually diminished according to the power law. CONCLUSIONS: We expanded and improved the existing DeepLesion dataset by annotating additional 21 775 lesions, and we demonstrated that using fully labeled CT images avoided overestimation of MVP-Net's performance while increasing the algorithm's sensitivity, which may have a huge impact to the future CT lesion detection research. The annotated lesions are at https://github.com/ComputationalImageAnalysisLab/DeepLesionData.

Machine learning-based identification of contrast-enhancement phase of computed tomography scans.

Contrast-enhanced computed tomography scans (CECT) are routinely used in the evaluation of different clinical scenarios, including the detection and characterization of hepatocellular carcinoma (HCC). Quantitative medical image analysis has been an exponentially growing scientific field. A number of studies reported on the effects of variations in the contrast enhancement phase on the reproducibility of quantitative imaging features extracted from CT scans. The identification and labeling of phase enhancement is a time-consuming task, with a current need for an accurate automated labeling algorithm to identify the enhancement phase of CT scans. In this study, we investigated the ability of machine learning algorithms to label the phases in a dataset of 59 HCC patients scanned with a dynamic contrast-enhanced CT protocol. The ground truth labels were provided by expert radiologists. Regions of interest were defined within the aorta, the portal vein, and the liver. Mean density values were extracted from those regions of interest and used for machine learning modeling. Models were evaluated using accuracy, the area under the curve (AUC), and Matthew's correlation coefficient (MCC). We tested the algorithms on an external dataset (76 patients). Our results indicate that several supervised learning algorithms (logistic regression, random forest, etc.) performed similarly, and our developed algorithms can accurately classify the phase of contrast enhancement.

Modeling Tumor Growth Using Partly Conditional Survival Models: A Case Study in Colorectal Cancer.

PURPOSE: There are multiple approaches to modeling the relationship between longitudinal tumor measurements obtained from serial imaging and overall survival. Many require strong assumptions that are untestable and debatable. We illustrate how to apply a novel, more flexible approach, the partly conditional (PC) survival model, using images acquired during a phase III, randomized clinical trial in colorectal cancer as an example. METHODS: PC survival approaches were used to model longitudinal volumetric computed tomography data of 1,025 patients in the completed VELOUR trial, which evaluated adding aflibercept to infusional fluorouracil, leucovorin, and irinotecan for treating metastatic colorectal cancer. PC survival modeling is a semiparametric approach to estimating associations of longitudinal measurements with time-to-event outcomes. Overall survival was our outcome. Covariates included baseline tumor burden, change in tumor burden from baseline to each follow-up time, and treatment. Both unstratified and time-stratified models were investigated. RESULTS: Without making assumptions about the distribution of the tumor growth process, we characterized associations between the change in tumor burden and survival. This change was significantly associated with survival (hazard ratio [HR], 1.04; 95% CI, 1.02 to 1.05; P < .001), suggesting that aflibercept works at least in part by altering the tumor growth trajectory. We also found baseline tumor size prognostic for survival even when accounting for the change in tumor burden over time (HR, 1.02; 95% CI, 1.01 to 1.02; P < .001). CONCLUSION: The PC modeling approach offers flexible characterization of associations between longitudinal covariates, such as serially assessed tumor burden, and survival time. It can be applied to a variety of data of this nature and used as clinical trials are ongoing to incorporate new disease assessment information as it is accumulated, as indicated by an example from colorectal cancer.

Ethnic diversity in treatment response for colorectal cancer: proof of concept for radiomics-driven enrichment trials.

BACKGROUND: Several barriers hamper recruitment of diverse patient populations in multicenter clinical trials which determine efficacy of new systemic cancer therapies. PURPOSE: We assessed if quantitative analysis of computed tomography (CT) scans of metastatic colorectal cancer (mCRC) patients using imaging features that predict overall survival (OS) can unravel the association between ethnicity and efficacy. METHODS: We retrospectively analyzed CT images from 1584 mCRC patients in two phase III trials evaluating FOLFOX ± panitumumab (n = 331, 350) and FOLFIRI ± aflibercept (n = 437, 466) collected from August 2006 to March 2013. Primary and secondary endpoints compared RECIST1.1 response at month-2 and delta tumor volume at month-2, respectively. An ancillary study compared imaging phenotype using a peer-reviewed radiomics-signature combining 3 imaging features to predict OS landmarked from month-2. Analysis was stratified by ethnicity. RESULTS: In total, 1584 patients were included (mean age, 60.25 ± 10.57 years; 969 men). Ethnicity was as follows: African (n = 50, 3.2%), Asian (n = 66, 4.2%), Caucasian (n = 1413, 89.2%), Latino (n = 27, 1.7%), Other (n = 28, 1.8%). Overall baseline tumor volume demonstrated Africans and Caucasians had more advanced disease (p < 0.001). Ethnicity was associated with treatment response. Response per RECIST1.1 at month-2 was distinct between ethnicities (p = 0.048) with higher response rate (55.6%) in Latinos. Overall delta tumor volume at month-2 demonstrated that Latino patients more likely experienced response to treatment (p = 0.021). Radiomics phenotype was also distinct in terms of tumor radiomics heterogeneity (p = 0.023). CONCLUSION: This study highlights how clinical trials that inadequately represent minority groups may impact associated translational work. In appropriately powered studies, radiomics features may allow us to unravel associations between ethnicity and treatment efficacy, better elucidate mechanisms of resistance, and promote diversity in trials through predictive enrichment. CLINICAL RELEVANCE STATEMENT: Radiomics could promote clinical trial diversity through predictive enrichment, hence benefit to historically underrepresented racial/ethnic groups that may respond variably to treatment due to socioeconomic factors and built environment, collectively referred to as social determinants of health. KEY POINTS: •Findings indicate ethnicity was associated with treatment response across all 3 endpoints. First, response per RECIST1.1 at month-2 was distinct between ethnicities (p = 0.048) with higher response rate (55.6%) in Latinos. •Second, the overall delta tumor volume at month-2 demonstrated that Latino patients were more likely to experience response to treatment (p = 0.021). Radiomics phenotype was also distinct in terms of tumor radiomics heterogeneity (p = 0.023).

Visceral adiposity in cirrhosis: Association with disease severity and impact of liver transplantation.

BACKGROUND: Changes in adipose tissue distribution in liver cirrhosis are poorly characterized and may affect clinical outcomes. METHODS: Adult liver transplant (LT) January 2008-August 2017 recipients with abdominal MRI within 6 months pre-LT were retrospectively assessed. Visceral adipose tissue, subcutaneous adipose tissue, and skeletal muscle area (cm2) were determined at L3. Visceral-to-subcutaneous adipose tissue ratio (VSR) was used to define relative adipose distribution, stratified by sex. Correlation was tested with Pearson. Body composition measures were compared by Child-Turcotte-Pugh (CTP) class, before and after LT, and evaluated as predictors of clinical outcomes. RESULTS: A total of 318 patients were studied. Mean age was 56 years, 33.64% were female, and 47.80% had CTP C cirrhosis. CTP C was associated with a 0.42-point increase in VSR compared with CTP A (95% CI = 0.13-0.71, p < 0.01), adjusting for age, sex, diabetes, and HCC. Among the 79 (24.84%) patients with repeat MRI 1-2 years after LT, VSR significantly improved from before LT (1.31 vs. 0.95, p < 0.01). In adjusted analysis, CTP C was associated with a 0.86-point decrease in post-LT VSR compared with pre-LT VSR (95% CI = -1.27 to -0.44, p < 0.01). Body mass index poorly correlated with VSR before and after LT. Elevated pre-LT VSR trended toward an association with a 7.17-point decrease in pre-LT glomerular filtration rate (95% CI = -14.35 to -0.02, p = 0.05), adjusting for CTP C, age, sex, diabetes, hypertension, pre-LT sarcopenia, and hepatocellular carcinoma. Elevated pre-LT VSR did not affect 3-year post-LT mortality (log-rank p = 0.24). CONCLUSIONS: Poorly represented by body mass index, visceral adiposity is increased in cirrhosis and is associated with CTP class. However, this adipose redistribution may be modifiable by LT.

Baseline Radiomic Signature to Estimate Overall Survival in Patients With NSCLC.

INTRODUCTION: We aimed to define a baseline radiomic signature associated with overall survival (OS) using baseline computed tomography (CT) images obtained from patients with NSCLC treated with nivolumab or chemotherapy. METHODS: The radiomic signature was developed in patients with NSCLC treated with nivolumab in CheckMate-017, -026, and -063. Nivolumab-treated patients were pooled and randomized to training, calibration, or validation sets using a 2:1:1 ratio. From baseline CT images, volume of tumor lesions was semiautomatically segmented, and 38 radiomic variables depicting tumor phenotype were extracted. Association between the radiomic signature and OS was assessed in the nivolumab-treated (validation set) and chemotherapy-treated (test set) patients in these studies. RESULTS: A baseline radiomic signature was identified using CT images obtained from 758 patients. The radiomic signature used a combination of imaging variables (spatial correlation, tumor volume in the liver, and tumor volume in the mediastinal lymph nodes) to output a continuous value, ranging from 0 to 1 (from most to least favorable estimated OS). Given a threshold of 0.55, the sensitivity and specificity of the radiomic signature for predicting 3-month OS were 86% and 77.8%, respectively. The signature was identified in the training set of patients treated with nivolumab and was significantly associated (p < 0.0001) with OS in patients treated with nivolumab or chemotherapy. CONCLUSIONS: The radiomic signature provides an early readout of the anticipated OS in patients with NSCLC treated with nivolumab or chemotherapy. This could provide important prognostic information and may support risk stratification in clinical trials.

Artificial intelligence and radiomics: fundamentals, applications, and challenges in immunotherapy.

Immunotherapy offers the potential for durable clinical benefit but calls into question the association between tumor size and outcome that currently forms the basis for imaging-guided treatment. Artificial intelligence (AI) and radiomics allow for discovery of novel patterns in medical images that can increase radiology's role in management of patients with cancer, although methodological issues in the literature limit its clinical application. Using keywords related to immunotherapy and radiomics, we performed a literature review of MEDLINE, CENTRAL, and Embase from database inception through February 2022. We removed all duplicates, non-English language reports, abstracts, reviews, editorials, perspectives, case reports, book chapters, and non-relevant studies. From the remaining articles, the following information was extracted: publication information, sample size, primary tumor site, imaging modality, primary and secondary study objectives, data collection strategy (retrospective vs prospective, single center vs multicenter), radiomic signature validation strategy, signature performance, and metrics for calculation of a Radiomics Quality Score (RQS). We identified 351 studies, of which 87 were unique reports relevant to our research question. The median (IQR) of cohort sizes was 101 (57-180). Primary stated goals for radiomics model development were prognostication (n=29, 33.3%), treatment response prediction (n=24, 27.6%), and characterization of tumor phenotype (n=14, 16.1%) or immune environment (n=13, 14.9%). Most studies were retrospective (n=75, 86.2%) and recruited patients from a single center (n=57, 65.5%). For studies with available information on model testing, most (n=54, 65.9%) used a validation set or better. Performance metrics were generally highest for radiomics signatures predicting treatment response or tumor phenotype, as opposed to immune environment and overall prognosis. Out of a possible maximum of 36 points, the median (IQR) of RQS was 12 (10-16). While a rapidly increasing number of promising results offer proof of concept that AI and radiomics could drive precision medicine approaches for a wide range of indications, standardizing the data collection as well as optimizing the methodological quality and rigor are necessary before these results can be translated into clinical practice.

Early Readout on Overall Survival of Patients With Melanoma Treated With Immunotherapy Using a Novel Imaging Analysis.

IMPORTANCE: Existing criteria to estimate the benefit of a therapy in patients with cancer rely almost exclusively on tumor size, an approach that was not designed to estimate survival benefit and is challenged by the unique properties of immunotherapy. More accurate prediction of survival by treatment could enhance treatment decisions. OBJECTIVE: To validate, using radiomics and machine learning, the performance of a signature of quantitative computed tomography (CT) imaging features for estimating overall survival (OS) in patients with advanced melanoma treated with immunotherapy. DESIGN, SETTING, AND PARTICIPANTS: This prognostic study used radiomics and machine learning to retrospectively analyze CT images obtained at baseline and first follow-up and their associated clinical metadata. Data were prospectively collected in the KEYNOTE-002 (Study of Pembrolizumab [MK-3475] Versus Chemotherapy in Participants With Advanced Melanoma; 2017 analysis) and KEYNOTE-006 (Study to Evaluate the Safety and Efficacy of Two Different Dosing Schedules of Pembrolizumab [MK-3475] Compared to Ipilimumab in Participants With Advanced Melanoma; 2016 analysis) multicenter clinical trials. Participants included 575 patients with a diagnosis of advanced melanoma who were randomly assigned to training and validation sets. Data for the present study were collected from November 20, 2012, to June 3, 2019, and analyzed from July 1, 2019, to September 15, 2021. INTERVENTIONS: KEYNOTE-002 featured trial groups testing intravenous pembrolizumab, 2 mg/kg or 10 mg/kg every 2 or every 3 weeks based on randomization, or investigator-choice chemotherapy; KEYNOTE-006 featured trial groups testing intravenous ipilimumab, 3 mg/kg every 3 weeks and intravenous pembrolizumab, 10 mg/kg every 2 or 3 weeks based on randomization. MAIN OUTCOMES AND MEASURES: The performance of the signature CT imaging features for estimating OS at the month 6 posttreatment landmark in patients who received pembrolizumab was measured using an area under the time-dependent receiver operating characteristics curve (AUC). RESULTS: A random forest model combined 25 imaging features extracted from tumors segmented on CT images to identify the combination (signature) that best estimated OS with pembrolizumab in 575 patients. The signature combined 4 imaging features, 2 related to tumor size and 2 reflecting changes in tumor imaging phenotype. In the validation set (287 patients treated with pembrolizumab), the signature reached an AUC for estimation of OS status of 0.92 (95% CI, 0.89-0.95). The standard method, Response Evaluation Criteria in Solid Tumors 1.1, achieved an AUC of 0.80 (95% CI, 0.75-0.84) and classified tumor outcomes as partial or complete response (93 of 287 [32.4%]), stable disease (90 of 287 [31.3%]), or progressive disease (104 of 287 [36.2%]). CONCLUSIONS AND RELEVANCE: The findings of this prognostic study suggest that the radiomic signature discerned from conventional CT images at baseline and on first follow-up may be used in clinical settings to provide an accurate early readout of future OS probability in patients with melanoma treated with single-agent programmed cell death 1 blockade.

Effect of CT image acquisition parameters on diagnostic performance of radiomics in predicting malignancy of pulmonary nodules of different sizes.

OBJECTIVES: To investigate the effect of CT image acquisition parameters on the performance of radiomics in classifying benign and malignant pulmonary nodules (PNs) with respect to nodule size. METHODS: We retrospectively collected CT images of 696 patients with PNs from March 2015 to March 2018. PNs were grouped by nodule diameter: T1a (diameter ≤ 1.0 cm), T1b (1.0 cm < diameter ≤ 2.0 cm), and T1c (2.0 cm < diameter ≤ 3.0 cm). CT images were divided into four settings according to slice-thickness-convolution-kernels: setting 1 (slice thickness/reconstruction type: 1.25 mm sharp), setting 2 (5 mm sharp), setting 3 (5 mm smooth), and random setting. We created twelve groups from two interacting conditions. Each PN was segmented and had 1160 radiomics features extracted. Non-redundant features with high predictive ability in training were selected to build a distinct model under each of the twelve subsets. RESULTS: The performance (AUCs) on predicting PN malignancy were as follows: T1a group: 0.84, 0.64, 0.68, and 0.68; T1b group: 0.68, 0.74, 0.76, and 0.70; T1c group: 0.66, 0.64, 0.63, and 0.70, for the setting 1, setting 2, setting 3, and random setting, respectively. In the T1a group, the AUC of radiomics model in setting 1 was statistically significantly higher than all others; In the T1b group, AUCs of radiomics models in setting 3 were statistically significantly higher than some; and in the T1c group, there were no statistically significant differences among models. CONCLUSIONS: For PNs less than 1 cm, CT image acquisition parameters have a significant influence on diagnostic performance of radiomics in predicting malignancy, and a model created using images reconstructed with thin section and a sharp kernel algorithm achieved the best performance. For PNs larger than 1 cm, CT reconstruction parameters did not affect diagnostic performance substantially. KEY POINTS: • CT image acquisition parameters have a significant influence on the diagnostic performance of radiomics in pulmonary nodules less than 1 cm. • In pulmonary nodules less than 1 cm, a radiomics model created by using images reconstructed with thin section and a sharp kernel algorithm achieved the best diagnostic performance. • For PNs larger than 1 cm, CT image acquisition parameters do not affect diagnostic performance substantially.

The Impact of Image Acquisition Parameters and ComBat Harmonization on the Predictive Performance of Radiomics: A Renal Cell Carcinoma Model.

Radiomics, one of the potential methods for developing clinical biomarker, is one of the exponentially growing research fields. In addition to its potential, several limitations have been identified in this field, and most importantly the effects of variations in imaging parameters on radiomic features (RFs). In this study, we investigate the potential of RFs to predict overall survival in patients with clear cell renal cell carcinoma, as well as the impact of ComBat harmonization on the performance of RF models. We assessed the robustness of the results by performing the analyses a thousand times. Publicly available CT scans of 179 patients were retrospectively collected and analyzed. The scans were acquired using different imaging vendors and parameters in different medical centers. The performance was calculated by averaging the metrics over all runs. On average, the clinical model significantly outperformed the radiomic models. The use of ComBat harmonization, on average, did not significantly improve the performance of radiomic models. Hence, the variability in image acquisition and reconstruction parameters significantly affect the performance of radiomic models. The development of radiomic specific harmonization techniques remain a necessity for the advancement of the field.

An imaging signature to predict outcome in metastatic colorectal cancer using routine computed tomography scans.

BACKGROUND & AIMS: Quantitative analysis of computed tomography (CT) scans of patients with metastatic colorectal cancer (mCRC) can identify imaging signatures that predict overall survival (OS). METHODS: We retrospectively analysed CT images from 1584 mCRC patients on two phase III trials evaluating FOLFOX ± panitumumab (n = 331, 350) and FOLFIRI ± aflibercept (n = 437, 466). In the training set (n = 720), an algorithm was trained to predict OS landmarked from month 2; the output was a signature value on a scale from 0 to 1 (most to least favourable predicted OS). In the validation set (n = 864), hazard ratios (HRs) evaluated the association of the signature with OS using RECIST1.1 as a benchmark of comparison. RESULTS: In the training set, the selected signature combined three features - change in tumour volume, change in tumour spatial heterogeneity, and tumour volume - to predict OS. In the validation set, RECIST1.1 classified patients in three categories: response (n = 166, 19.2%), stable disease (n = 636, 73.6%), and progression (n = 62, 7.2%). The HR was 3.93 (2.79-5.54). Using the same distribution for the signature, the HR was 21.04 (14.88-30.58), showing an incremental prognostic separation. Stable disease by RECIST1.1 was reclassified by the signature along a continuum where patients belonging to the most and least favourable signature quartiles had a median OS of 40.73 (28.49 to NA) months (n = 94) and 7.03 (5.66-7.89) months (n = 166), respectively. CONCLUSIONS: A signature combining three imaging features provides early prognostic information that can improve treatment decisions for individual patients and clinical trial analyses.

Convolutional Neural Network Addresses the Confounding Impact of CT Reconstruction Kernels on Radiomics Studies.

Achieving high feature reproducibility while preserving biological information is one of the main challenges for the generalizability of current radiomics studies. Non-clinical imaging variables, such as reconstruction kernels, have shown to significantly impact radiomics features. In this study, we retrain an open-source convolutional neural network (CNN) to harmonize computerized tomography (CT) images with various reconstruction kernels to improve feature reproducibility and radiomic model performance using epidermal growth factor receptor (EGFR) mutation prediction in lung cancer as a paradigm. In the training phase, the CNN was retrained and tested on 32 lung cancer patients' CT images between two different groups of reconstruction kernels (smooth and sharp). In the validation phase, the retrained CNN was validated on an external cohort of 223 lung cancer patients' CT images acquired using different CT scanners and kernels. The results showed that the retrained CNN could be successfully applied to external datasets with different CT scanner parameters, and harmonization of reconstruction kernels from sharp to smooth could significantly improve the performance of radiomics model in predicting EGFR mutation status in lung cancer. In conclusion, the CNN based method showed great potential in improving feature reproducibility and generalizability by harmonizing medical images with heterogeneous reconstruction kernels.

Deep learning for the prediction of early on-treatment response in metastatic colorectal cancer from serial medical imaging.

In current clinical practice, tumor response assessment is usually based on tumor size change on serial computerized tomography (CT) scan images. However, evaluation of tumor response to anti-vascular endothelial growth factor therapies in metastatic colorectal cancer (mCRC) is limited because morphological change in tumor may occur earlier than tumor size change. Here we present an analysis utilizing a deep learning (DL) network to characterize tumor morphological change for response assessment in mCRC patients. We retrospectively analyzed 1,028 mCRC patients who were prospectively included in the VELOUR trial (NCT00561470). We found that DL network was able to predict early on-treatment response in mCRC and showed better performance than its size-based counterpart with C-Index: 0.649 (95% CI: 0.619,0.679) vs. 0.627 (95% CI: 0.567,0.638), p = 0.009, z-test. The integration of DL network with size-based methodology could further improve the prediction performance to C-Index: 0.694 (95% CI: 0.661,0.720), which was superior to size/DL-based-only models (all p < 0.001, z-test). Our study suggests that DL network could provide a noninvasive mean for quantitative and comprehensive characterization of tumor morphological change, which may potentially benefit personalized early on-treatment decision making.

Predicting Tyrosine Kinase Inhibitor Treatment Response in Stage IV Lung Adenocarcinoma Patients With EGFR Mutation Using Model-Based Deep Transfer Learning.

BACKGROUND: For stage IV patients harboring EGFR mutations, there is a differential response to the first-line TKI treatment. We constructed three-dimensional convolutional neural networks (CNN) with deep transfer learning to stratify patients into subgroups with different response and progression risks. MATERIALS AND METHODS: From 2013 to 2017, 339 patients with EGFR mutation receiving first-line TKI treatment were included. Progression-free survival (PFS) time and progression patterns were confirmed by routine follow-up and restaging examinations. Patients were divided into two subgroups according to the median PFS (<=9 months, > 9 months). We developed a PFS prediction model and a progression pattern classification model using transfer learning from a pre-trained EGFR mutation classification 3D CNN. Clinical features were fused with the 3D CNN to build the final hybrid prediction model. The performance was quantified using area under receiver operating characteristic curve (AUC), and model performance was compared by AUCs with Delong test. RESULTS: The PFS prediction CNN showed an AUC of 0.744 (95% CI, 0.645-0.843) in the independent validation set and the hybrid model of CNNs and clinical features showed an AUC of 0.771 (95% CI, 0.676-0.866), which are significantly better than clinical features-based model (AUC, 0.624, P<0.01). The progression pattern prediction model showed an AUC of 0.762(95% CI, 0.643-0.882) and the hybrid model with clinical features showed an AUC of 0.794 (95% CI, 0.681-0.908), which can provide compensate information for clinical features-based model (AUC, 0.710; 95% CI, 0.582-0.839). CONCLUSION: The CNN exhibits potential ability to stratify progression status in patients with EGFR mutation treated with first-line TKI, which might help make clinical decisions.

Distinguishing benign and malignant lesions on contrast-enhanced breast cone-beam CT with deep learning neural architecture search.

PURPOSE: To utilize a neural architecture search (NAS) approach to develop a convolutional neural network (CNN) method for distinguishing benign and malignant lesions on breast cone-beam CT (BCBCT). METHOD: 165 patients with 114 malignant and 86 benign lesions were collected by two institutions from May 2012 to August 2014. The NAS method autonomously generated a CNN model using one institution's dataset for training (patients/lesions: 71/91) and validation (patients/lesions: 20/23). The model was externally tested on another institution's dataset (patients/lesions: 74/87), and its performance was compared with fine-tuned ResNet-50 models and two breast radiologists who independently read the lesions in the testing dataset without knowing lesion diagnosis. RESULTS: The lesion diameters (mean ± SD) were 18.8 ± 12.9 mm, 22.7 ± 10.5 mm, and 20.0 ± 11.8 mm in the training, validation, and external testing set, respectively. Compared to the best ResNet-50 model, the NAS-generated CNN model performed three times faster and, in the external testing set, achieved a higher (though not statistically different) AUC, with sensitivity (95% CI) and specificity (95% CI) of 0.727, 80% (66-90%), and 60% (42-75%), respectively. Meanwhile, the performances of the NAS-generated CNN and the two radiologists' visual ratings were not statistically different. CONCLUSIONS: Our preliminary results demonstrated that a CNN autonomously generated by NAS performed comparably to pre-trained ResNet models and radiologists in predicting malignant breast lesions on contrast-enhanced BCBCT. In comparison to ResNet, which must be designed by an expert, the NAS approach may be used to automatically generate a deep learning architecture for medical image analysis.

Uncontrolled Confounders May Lead to False or Overvalued Radiomics Signature: A Proof of Concept Using Survival Analysis in a Multicenter Cohort of Kidney Cancer.

PURPOSE: We aimed to explore potential confounders of prognostic radiomics signature predicting survival outcomes in clear cell renal cell carcinoma (ccRCC) patients and demonstrate how to control for them. MATERIALS AND METHODS: Preoperative contrast enhanced abdominal CT scan of ccRCC patients along with pathological grade/stage, gene mutation status, and survival outcomes were retrieved from The Cancer Imaging Archive (TCIA)/The Cancer Genome Atlas-Kidney Renal Clear Cell Carcinoma (TCGA-KIRC) database, a publicly available dataset. A semi-automatic segmentation method was applied to segment ccRCC tumors, and 1,160 radiomics features were extracted from each segmented tumor on the CT images. Non-parametric principal component decomposition (PCD) and unsupervised hierarchical clustering were applied to build the radiomics signature models. The factors confounding the radiomics signature were investigated and controlled sequentially. Kaplan-Meier curves and Cox regression analyses were performed to test the association between radiomics signatures and survival outcomes. RESULTS: 183 patients of TCGA-KIRC cohort with available imaging, pathological, and clinical outcomes were included in this study. All 1,160 radiomics features were included in the first radiomics signature. Three additional radiomics signatures were then modelled in successive steps removing redundant radiomics features first, removing radiomics features biased by CT slice thickness second, and removing radiomics features dependent on tumor size third. The final radiomics signature model was the most parsimonious, unbiased by CT slice thickness, and independent of tumor size. This final radiomics signature stratified the cohort into radiomics phenotypes that are different by cancer-specific and recurrence-free survival; HR (95% CI) = 3.0 (1.5-5.7), p <0.05 and HR (95% CI) = 6.6 (3.1-14.1), p <0.05, respectively. CONCLUSION: Radiomics signature can be confounded by multiple factors, including feature redundancy, image acquisition parameters like slice thickness, and tumor size. Attention to and proper control for these potential confounders are necessary for a reliable and clinically valuable radiomics signature.

Understanding Sources of Variation to Improve the Reproducibility of Radiomics.

Radiomics is the method of choice for investigating the association between cancer imaging phenotype, cancer genotype and clinical outcome prediction in the era of precision medicine. The fast dispersal of this new methodology has benefited from the existing advances of the core technologies involved in radiomics workflow: image acquisition, tumor segmentation, feature extraction and machine learning. However, despite the rapidly increasing body of publications, there is no real clinical use of a developed radiomics signature so far. Reasons are multifaceted. One of the major challenges is the lack of reproducibility and generalizability of the reported radiomics signatures (features and models). Sources of variation exist in each step of the workflow; some are controllable or can be controlled to certain degrees, while others are uncontrollable or even unknown. Insufficient transparency in reporting radiomics studies further prevents translation of the developed radiomics signatures from the bench to the bedside. This review article first addresses sources of variation, which is illustrated using demonstrative examples. Then, it reviews a number of published studies and progresses made to date in the investigation and improvement of feature reproducibility and model performance. Lastly, it discusses potential strategies and practical considerations to reduce feature variability and improve the quality of radiomics study. This review focuses on CT image acquisition, tumor segmentation, quantitative feature extraction, and the disease of lung cancer.

Identifying Robust Radiomics Features for Lung Cancer by Using In-Vivo and Phantom Lung Lesions.

We propose a novel framework for determining radiomics feature robustness by considering the effects of both biological and noise signals. This framework is preliminarily tested in a study predicting the epidermal growth factor receptor (EGFR) mutation status in non-small cell lung cancer (NSCLC) patients. Pairs of CT images (baseline, 3-week post therapy) of 46 NSCLC patients with known EGFR mutation status were collected and a FDA-customized anthropomorphic thoracic phantom was scanned on two vendors' scanners at four different tube currents. Delta radiomics features were extracted from the NSCLC patient CTs and reproducible, non-redundant, and informative features were identified. The feature value differences between EGFR mutant and EGFR wildtype patients were quantitatively measured as the biological signal. Similarly, radiomics features were extracted from the phantom CTs. A pairwise comparison between settings resulted in a feature value difference that was quantitatively measured as the noise signal. Biological signals were compared to noise signals at each setting to determine if the distributions were significantly different by two-sample t-test, and thus robust. Four optimal features were selected to predict EGFR mutation status, Tumor-Mass, Sigmoid-Offset-Mean, Gabor-Energy and DWT-Energy, which quantified tumor mass, tumor-parenchyma density transition at boundary, line-like pattern inside tumor and intratumoral heterogeneity, respectively. The first three variables showed robustness across the majority of studied CT acquisition parameters. The textual feature DWT-Energy was less robust. The proposed framework was able to determine robustness of radiomics features at specific settings by comparing biological signal to noise signal. Identification of robust radiomics features may improve the generalizability of radiomics models in future studies.

Comparing RECIST 1.1 and iRECIST in advanced melanoma patients treated with pembrolizumab in a phase II clinical trial.

OBJECTIVES: To compare tumor best overall response (BOR) by RECIST 1.1 and iRECIST, to explore the incidence of pseudoprogression in melanoma treated with pembrolizumab, and to assess the impact of pseudoprogression on overall survival (OS). METHODS: A total of 221 patients with locally advanced/unresectable melanoma who received pembrolizumab as part of KEYNOTE-002 trial were included in this study. Radiological assessment of imaging was centrally reviewed to assess tumor response. Incidence of discordance in BOR between RECIST 1.1 and iRECIST as well as rate of pseudoprogression were measured. OS of patients with pseudoprogression was compared with that of those with uncontrolled disease. RESULTS: Of the 221 patients in this cohort, 136 patients developed PD as per RECIST v1.1 and 78 patients with PD continued treatment and imaging beyond initial RECIST 1.1-defined PD. Among the 78 patients who continued therapy and imaging post-progression, RECIST 1.1 and iRECIST were discordant in 10 patients (12.8%) and pseudoprogression was encountered in 14 patients (17.9%). OS of patients with pseudoprogression was longer than that of patients with uncontrolled disease/true progression (29.9 months versus 8.0 months, p value < 0.001). CONCLUSIONS: Effectiveness of immunotherapy in clinical trials depends on the criterion used to assess tumor response (RECIST 1.1 vs iRECIST) with iRECIST being more appropriate to detect pseudoprogression and potentially prevent premature termination of effective therapy. Pseudoprogression was associated with improved OS in comparison with that of patients with uncontrolled disease. KEY POINTS: • Discordance between iRECIST and RECIST 1.1 was found in 12.8% of unresectable melanoma patients on pembrolizumab who continued therapy beyond initial RECIST 1.1-defined progression. • Pseudoprogression, captured with iRECIST, occurred in 17.9% and was significantly associated with improved overall survival in comparison with uncontrolled disease.