Structural and functional radiomics for lung cancer.

INTRODUCTION: Lung cancer ranks second in new cancer cases and first in cancer-related deaths worldwide. Precision medicine is working on altering treatment approaches and improving outcomes in this patient population. Radiological images are a powerful non-invasive tool in the screening and diagnosis of early-stage lung cancer, treatment strategy support, prognosis assessment, and follow-up for advanced-stage lung cancer. Recently, radiological features have evolved from solely semantic to include (handcrafted and deep) radiomic features. Radiomics entails the extraction and analysis of quantitative features from medical images using mathematical and machine learning methods to explore possible ties with biology and clinical outcomes. METHODS: Here, we outline the latest applications of both structural and functional radiomics in detection, diagnosis, and prediction of pathology, gene mutation, treatment strategy, follow-up, treatment response evaluation, and prognosis in the field of lung cancer. CONCLUSION: The major drawbacks of radiomics are the lack of large datasets with high-quality data, standardization of methodology, the black-box nature of deep learning, and reproducibility. The prerequisite for the clinical implementation of radiomics is that these limitations are addressed. Future directions include a safer and more efficient model-training mode, merge multi-modality images, and combined multi-discipline or multi-omics to form "Medomics."

Preoperative CT-based radiomics combined with intraoperative frozen section is predictive of invasive adenocarcinoma in pulmonary nodules: a multicenter study.

OBJECTIVES: Develop a CT-based radiomics model and combine it with frozen section (FS) and clinical data to distinguish invasive adenocarcinomas (IA) from preinvasive lesions/minimally invasive adenocarcinomas (PM). METHODS: This multicenter study cohort of 623 lung adenocarcinomas was split into training (n = 331), testing (n = 143), and external validation dataset (n = 149). Random forest models were built using selected radiomics features, results from FS, lesion volume, clinical and semantic features, and combinations thereof. The area under the receiver operator characteristic curves (AUC) was used to evaluate model performances. The diagnosis accuracy, calibration, and decision curves of models were tested. RESULTS: The radiomics-based model shows good predictive performance and diagnostic accuracy for distinguishing IA from PM, with AUCs of 0.89, 0.89, and 0.88, in the training, testing, and validation datasets, respectively, and with corresponding accuracies of 0.82, 0.79, and 0.85. Adding lesion volume and FS significantly increases the performance of the model with AUCs of 0.96, 0.97, and 0.96, and with accuracies of 0.91, 0.94, and 0.93 in the three datasets. There is no significant difference in AUC between the FS model enriched with radiomics and volume against an FS model enriched with volume alone, while the former has higher accuracy. The model combining all available information shows minor non-significant improvements in AUC and accuracy compared with an FS model enriched with radiomics and volume. CONCLUSIONS: Radiomics signatures are potential biomarkers for the risk of IA, especially in combination with FS, and could help guide surgical strategy for pulmonary nodules patients. KEY POINTS: • A CT-based radiomics model may be a valuable tool for preoperative prediction of invasive adenocarcinoma for patients with pulmonary nodules. • Radiomics combined with frozen sections could help in guiding surgery strategy for patients with pulmonary nodules.

Deep learning in fracture detection: a narrative review.

Artificial intelligence (AI) is a general term that implies the use of a computer to model intelligent behavior with minimal human intervention. AI, particularly deep learning, has recently made substantial strides in perception tasks allowing machines to better represent and interpret complex data. Deep learning is a subset of AI represented by the combination of artificial neuron layers. In the last years, deep learning has gained great momentum. In the field of orthopaedics and traumatology, some studies have been done using deep learning to detect fractures in radiographs. Deep learning studies to detect and classify fractures on computed tomography (CT) scans are even more limited. In this narrative review, we provide a brief overview of deep learning technology: we (1) describe the ways in which deep learning until now has been applied to fracture detection on radiographs and CT examinations; (2) discuss what value deep learning offers to this field; and finally (3) comment on future directions of this technology.

Blood steroids are associated with prognosis and fat distribution in endometrial cancer.

BACKGROUND: Despite being a hormone dependent cancer, there is limited knowledge regarding the relation between level of steroids in blood and prognosis for endometrial cancer (EC) patients. METHODS: In this study we investigated plasma levels of 19 steroids using liquid-chromatography tandem mass-spectrometry in 38 postmenopausal EC patients, 19 with long, and 19 with short survival. We explored if estradiol levels were associated with specific abdominal fat distribution patterns and if transcriptional alterations related to estradiol levels could be observed in tumor samples. RESULTS: The plasma steroid levels for DHEA, DHEAS, progesterone, 21 OH progesterone and E1S were significantly increased (all p < 0.05) in patients with long survival compared to short. Estradiol levels were significantly positively correlated with visceral fat percentage (p = 0.035), and an increased expression of genes involved in estrogen related signaling was observed in tumors from patients with high estradiol levels in plasma. CONCLUSION: Several of the identified plasma steroids represent promising biomarkers in EC patients. The association between increased estradiol levels and a high percentage of visceral fat indicates that visceral fat is a larger contributor to estradiol production compared to subcutaneous fat in this population.

Development of an isotoxic decision support system integrating genetic markers of toxicity for the implantation of a rectum spacer.

INTRODUCTION: Previous studies revealed that dose escalated radiotherapy for prostate cancer patients leads to higher tumor control probabilities (TCP) but also to higher rectal toxicities. An isotoxic model was developed to maximize the given dose while controlling the toxicity level. This was applied to analyze the effect of an implantable rectum spacer (IRS) and extended with a genetic test of normal tissue radio-sensitivity. A virtual IRS (V-IRS) was tested using this method. We hypothesized that the patients with increased risk of toxicity would benefit more from an IRS. MATERIAL AND METHODS: Sixteen localized prostate cancer patients implanted with an IRS were included in the study. Treatment planning was performed on computed tomography (CT) images before and after the placement of the IRS and with a V-IRS. The normal tissue complication probability (NTCP) was calculated using a QUANTEC reviewed model for Grade > =2 late rectal bleeding and the number of fractions of the plans were adjusted until the NTCP value was under 5%. The resulting treatment plans were used to calculate the TCP before and after placement of an IRS. This was extended by adding the effect of two published genetic single nucleotide polymorphisms (SNP's) for late rectal bleeding. RESULTS: The median TCP resulting from the optimized plans in patients before the IRS was 75.1% [32.6-90.5%]. With IRS, the median TCP is significantly higher: 98.9% [80.8-99.9%] (p < .01). The difference in TCP between the V-IRS and the real IRS was 1.8% [0.0-18.0%]. Placing an IRS in the patients with SNP's improved the TCP from 49.0% [16.1-80.8%] and 48.9% [16.0-72.8%] to 96.3% [67.0-99.5%] and 90.1% [49.0-99.5%] (p < .01) respectively for either SNP. CONCLUSION: This study was a proof-of-concept for an isotoxic model with genetic biomarkers with a V-IRS as a multifactorial decision support system for the decision of a placement of an IRS.

Pre-treatment CT radiomics to predict 3-year overall survival following chemoradiotherapy of esophageal cancer.

BACKGROUND: Radiomic features retrieved from standard CT-images have shown prognostic power in several tumor sites. In this study, we investigated the prognostic value of pretreatment CT radiomic features to predict overall survival of esophageal cancer patients after chemoradiotherapy. MATERIAL AND METHODS: Two datasets of independent centers were analyzed, consisting of esophageal cancer patients treated with concurrent chemotherapy (Carboplatin/Paclitaxel) and 41.4Gy radiotherapy, followed by surgery if feasible. In total, 1049 radiomic features were calculated from the primary tumor volume. Recursive feature elimination was performed to select the 40 most relevant predictors. Using these 40 features and six clinical variables as input, two random forest (RF) models predicting 3-year overall survival were developed. RESULTS: In total 165 patients from center 1 and 74 patients from center 2 were used. The radiomics-based RF model yielded an area under the curve (AUC) of 0.69 (95%CI 0.61-0.77), with the top-5 most important features for 3-year survival describing tumor heterogeneity after wavelet filtering. In the validation dataset, the RF model yielded an AUC of 0.61 (95%CI 0.47-0.75). Kaplan Meier plots were significantly different between risk groups in the training dataset (p = .027) and borderline significant in the validation dataset (p = .053). The clinical RF model yielded AUCs of 0.63 (95%CI 0.54-0.71) and 0.62 (95%CI 0.49-0.76) in the training and validation dataset, respectively. Risk groups did not reach a significant correlation with pathological response in the primary tumor. CONCLUSIONS: A RF model predicting 3-year overall survival based on pretreatment CT radiomic features was developed and validated in two independent datasets of esophageal cancer patients. The radiomics model had better prognostic power compared to the model using standard clinical variables.

A prediction model for early death in non-small cell lung cancer patients following curative-intent chemoradiotherapy.

BACKGROUND: Early death after a treatment can be seen as a therapeutic failure. Accurate prediction of patients at risk for early mortality is crucial to avoid unnecessary harm and reducing costs. The goal of our work is two-fold: first, to evaluate the performance of a previously published model for early death in our cohorts. Second, to develop a prognostic model for early death prediction following radiotherapy. MATERIAL AND METHODS: Patients with NSCLC treated with chemoradiotherapy or radiotherapy alone were included in this study. Four different cohorts from different countries were available for this work (N = 1540). The previous model used age, gender, performance status, tumor stage, income deprivation, no previous treatment given (yes/no) and body mass index to make predictions. A random forest model was developed by learning on the Maastro cohort (N = 698). The new model used performance status, age, gender, T and N stage, total tumor volume (cc), total tumor dose (Gy) and chemotherapy timing (none, sequential, concurrent) to make predictions. Death within 4 months of receiving the first radiotherapy fraction was used as the outcome. RESULTS: Early death rates ranged from 6 to 11% within the four cohorts. The previous model performed with AUC values ranging from 0.54 to 0.64 on the validation cohorts. Our newly developed model had improved AUC values ranging from 0.62 to 0.71 on the validation cohorts. CONCLUSIONS: Using advanced machine learning methods and informative variables, prognostic models for early mortality can be developed. Development of accurate prognostic tools for early mortality is important to inform patients about treatment options and optimize care.

Predicting tumor hypoxia in non-small cell lung cancer by combining CT, FDG PET and dynamic contrast-enhanced CT.

BACKGROUND: Most solid tumors contain inadequately oxygenated (i.e., hypoxic) regions, which tend to be more aggressive and treatment resistant. Hypoxia PET allows visualization of hypoxia and may enable treatment adaptation. However, hypoxia PET imaging is expensive, time-consuming and not widely available. We aimed to predict hypoxia levels in non-small cell lung cancer (NSCLC) using more easily available imaging modalities: FDG-PET/CT and dynamic contrast-enhanced CT (DCE-CT). MATERIAL AND METHODS: For 34 NSCLC patients, included in two clinical trials, hypoxia HX4-PET/CT, planning FDG-PET/CT and DCE-CT scans were acquired before radiotherapy. Scans were non-rigidly registered to the planning CT. Tumor blood flow (BF) and blood volume (BV) were calculated by kinetic analysis of DCE-CT images. Within the gross tumor volume, independent clusters, i.e., supervoxels, were created based on FDG-PET/CT. For each supervoxel, tumor-to-background ratios (TBR) were calculated (median SUV/aorta SUVmean) for HX4-PET/CT and supervoxel features (median, SD, entropy) for the other modalities. Two random forest models (cross-validated: 10 folds, five repeats) were trained to predict the hypoxia TBR; one based on CT, FDG, BF and BV, and one with only CT and FDG features. Patients were split in a training (trial NCT01024829) and independent test set (trial NCT01210378). For each patient, predicted, and observed hypoxic volumes (HV) (TBR > 1.2) were compared. RESULTS: Fifteen patients (3291 supervoxels) were used for training and 19 patients (1502 supervoxels) for testing. The model with all features (RMSE training: 0.19 ± 0.01, test: 0.27) outperformed the model with only CT and FDG-PET features (RMSE training: 0.20 ± 0.01, test: 0.29). All tumors of the test set were correctly classified as normoxic or hypoxic (HV > 1 cm3) by the best performing model. CONCLUSIONS: We created a data-driven methodology to predict hypoxia levels and hypoxia spatial patterns using CT, FDG-PET and DCE-CT features in NSCLC. The model correctly classifies all tumors, and could therefore, aid tumor hypoxia classification and patient stratification.

Transition between encoding and consolidation/replay dynamics via cholinergic modulation of CAN current: A modeling study.

Hippocampal place cells that are activated sequentially during active waking get reactivated in a temporally compressed (5-20 times) manner during slow-wave-sleep and quiet waking. The two-stage model of the hippocampus suggests that neural activity during awaking supports encoding function while temporally compressed reactivation (replay) supports consolidation. However, the mechanisms supporting different neural activity with different temporal scales during encoding and consolidation remain unclear. Based on the idea that acetylcholine modulates functional transition between encoding and consolidation, we tested whether the cholinergic modulation may adjust intrinsic network dynamics to support different temporal scales for these two modes of operation. Simulations demonstrate that cholinergic modulation of the calcium activated non-specific cationic (CAN) current and the synaptic transmission may be sufficient to switch the network dynamics between encoding and consolidation modes. When the CAN current is active and the synaptic transmission is suppressed, mimicking the high acetylcholine condition during active waking, a slow propagation of multiple spikes is evident. This activity resembles the firing pattern of place cells and time cells during active waking. On the other hand, when CAN current is suppressed and the synaptic transmission is intact, mimicking the low acetylcholine condition during slow-wave-sleep, a time compressed fast (∼10 times) activity propagation of the same set of cells is evident. This activity resembles the time compressed firing pattern of place cells during replay and pre-play, achieving a temporal compression factor in the range observed in vivo (5-20 times). These observations suggest that cholinergic system could adjust intrinsic network dynamics suitable for encoding and consolidation through the modulation of the CAN current and synaptic conductance in the hippocampus.

Long-lasting intrinsic persistent firing in rat CA1 pyramidal cells: a possible mechanism for active maintenance of memory.

The hippocampus is critical for memory tasks which require an active maintenance of memory for a short period of time; however, the underlying neural mechanisms remain unknown. Most theoretical and computational models, which date back to the classic proposals by Donald Hebb in , have been self-constrained by anatomy, as most models rely on the recurrent connectivity in region CA3 to support "reverberating activity" capable of memory maintenance. However, several physiological and behavioral studies have specifically implicated region CA1 in tasks which require an active maintenance of memory. Here, we demonstrate that despite limited recurrent connectivity, CA1 contains a robust cellular mechanism for active memory maintenance in the form of self-sustained persistent firing. Using in vitro whole-cell recordings, we demonstrate that brief stimulation (0.2-2 s) reliably elicits long-lasting (> 30 s) persistent firing that is supported by the calcium-activated non-selective cationic current. In contrast to more traditional ideas, these data suggest that the hippocampal region CA1 is capable of active maintenance of memory.

Persistent firing supported by an intrinsic cellular mechanism in hippocampal CA3 pyramidal cells.

Short-term information retention is crucial for information processing in the brain. It has long been suggested that the hippocampal CA3 region is able to support short-term information retention through persistent neural firing. Theoretical studies have shown that this persistent firing can be supported by abundant excitatory recurrent connections in CA3. However, it remains unclear whether individual cells can support persistent firing. In this study, using in vitro whole-cell patch-clamp recordings in a rat hippocampal slice preparation, we show that hippocampal CA3 pyramidal cells support persistent firing under perfusion of the cholinergic agonist carbachol (10 µm). Furthermore, in contrast to earlier theoretical studies, this persistent firing is independent of ionotropic glutamatergic synaptic transmission and is supported by the calcium-activated non-selective cationic current. Because cholinergic receptor activation is crucial for short-term memory tasks, persistent firing in individual cells may support short-term information retention in the hippocampal CA3 region.

Improving shared decision making for lung cancer treatment by developing and validating an open-source web based patient decision aid for stage I-II non-small cell lung cancer.

The aim of this study was to develop and evaluate a proof-of-concept open-source individualized Patient Decision Aid (iPDA) with a group of patients, physicians, and computer scientists. The iPDA was developed based on the International Patient Decision Aid Standards (IPDAS). A previously published questionnaire was adapted and used to test the user-friendliness and content of the iPDA. The questionnaire contained 40 multiple-choice questions, and answers were given on a 5-point Likert Scale (1-5) ranging from "strongly disagree" to "strongly agree." In addition to the questionnaire, semi-structured interviews were conducted with patients. We performed a descriptive analysis of the responses. The iPDA was evaluated by 28 computer scientists, 21 physicians, and 13 patients. The results demonstrate that the iPDA was found valuable by 92% (patients), 96% (computer scientists), and 86% (physicians), while the treatment information was judged useful by 92%, 96%, and 95%, respectively. Additionally, the tool was thought to be motivating for patients to actively engage in their treatment by 92%, 93%, and 91% of the above respondents groups. More multimedia components and less text were suggested by the respondents as ways to improve the tool and user interface. In conclusion, we successfully developed and tested an iPDA for patients with stage I-II Non-Small Cell Lung Cancer (NSCLC).

Automated detection and segmentation of non-small cell lung cancer computed tomography images.

Detection and segmentation of abnormalities on medical images is highly important for patient management including diagnosis, radiotherapy, response evaluation, as well as for quantitative image research. We present a fully automated pipeline for the detection and volumetric segmentation of non-small cell lung cancer (NSCLC) developed and validated on 1328 thoracic CT scans from 8 institutions. Along with quantitative performance detailed by image slice thickness, tumor size, image interpretation difficulty, and tumor location, we report an in-silico prospective clinical trial, where we show that the proposed method is faster and more reproducible compared to the experts. Moreover, we demonstrate that on average, radiologists & radiation oncologists preferred automatic segmentations in 56% of the cases. Additionally, we evaluate the prognostic power of the automatic contours by applying RECIST criteria and measuring the tumor volumes. Segmentations by our method stratified patients into low and high survival groups with higher significance compared to those methods based on manual contours.

Modeling-Based Decision Support System for Radical Prostatectomy Versus External Beam Radiotherapy for Prostate Cancer Incorporating an In Silico Clinical Trial and a Cost-Utility Study.

The aim of this study is to build a decision support system (DSS) to select radical prostatectomy (RP) or external beam radiotherapy (EBRT) for low- to intermediate-risk prostate cancer patients. We used an individual state-transition model based on predictive models for estimating tumor control and toxicity probabilities. We performed analyses on a synthetically generated dataset of 1000 patients with realistic clinical parameters, externally validated by comparison to randomized clinical trials, and set up an in silico clinical trial for elderly patients. We assessed the cost-effectiveness (CE) of the DSS for treatment selection by comparing it to randomized treatment allotment. Using the DSS, 47.8% of synthetic patients were selected for RP and 52.2% for EBRT. During validation, differences with the simulations of late toxicity and biochemical failure never exceeded 2%. The in silico trial showed that for elderly patients, toxicity has more influence on the decision than TCP, and the predicted QoL depends on the initial erectile function. The DSS is estimated to result in cost savings (EUR 323 (95% CI: EUR 213-433)) and more quality-adjusted life years (QALYs; 0.11 years, 95% CI: 0.00-0.22) than randomized treatment selection.

Selectively Targeting Tumor Hypoxia With the Hypoxia-Activated Prodrug CP-506.

Hypoxia-activated prodrugs (HAP) are a promising class of antineoplastic agents that can selectively eliminate hypoxic tumor cells. This study evaluates the hypoxia-selectivity and antitumor activity of CP-506, a DNA alkylating HAP with favorable pharmacologic properties. Stoichiometry of reduction, one-electron affinity, and back-oxidation rate of CP-506 were characterized by fast-reaction radiolytic methods with observed parameters fulfilling requirements for oxygen-sensitive bioactivation. Net reduction, metabolism, and cytotoxicity of CP-506 were maximally inhibited at oxygen concentrations above 1 µmol/L (0.1% O2). CP-506 demonstrated cytotoxicity selectively in hypoxic 2D and 3D cell cultures with normoxic/anoxic IC50 ratios up to 203. Complete resistance to aerobic (two-electron) metabolism by aldo-keto reductase 1C3 was confirmed through gain-of-function studies while retention of hypoxic (one-electron) bioactivation by various diflavin oxidoreductases was also demonstrated. In vivo, the antitumor effects of CP-506 were selective for hypoxic tumor cells and causally related to tumor oxygenation. CP-506 effectively decreased the hypoxic fraction and inhibited growth of a wide range of hypoxic xenografts. A multivariate regression analysis revealed baseline tumor hypoxia and in vitro sensitivity to CP-506 were significantly correlated with treatment response. Our results demonstrate that CP-506 selectively targets hypoxic tumor cells and has broad antitumor activity. Our data indicate that tumor hypoxia and cellular sensitivity to CP-506 are strong determinants of the antitumor effects of CP-506.

Systematic Review of Privacy-Preserving Distributed Machine Learning From Federated Databases in Health Care.

Big data for health care is one of the potential solutions to deal with the numerous challenges of health care, such as rising cost, aging population, precision medicine, universal health coverage, and the increase of noncommunicable diseases. However, data centralization for big data raises privacy and regulatory concerns.Covered topics include (1) an introduction to privacy of patient data and distributed learning as a potential solution to preserving these data, a description of the legal context for patient data research, and a definition of machine/deep learning concepts; (2) a presentation of the adopted review protocol; (3) a presentation of the search results; and (4) a discussion of the findings, limitations of the review, and future perspectives.Distributed learning from federated databases makes data centralization unnecessary. Distributed algorithms iteratively analyze separate databases, essentially sharing research questions and answers between databases instead of sharing the data. In other words, one can learn from separate and isolated datasets without patient data ever leaving the individual clinical institutes.Distributed learning promises great potential to facilitate big data for medical application, in particular for international consortiums. Our purpose is to review the major implementations of distributed learning in health care.

Distributed learning on 20 000+ lung cancer patients - The Personal Health Train.

BACKGROUND AND PURPOSE: Access to healthcare data is indispensable for scientific progress and innovation. Sharing healthcare data is time-consuming and notoriously difficult due to privacy and regulatory concerns. The Personal Health Train (PHT) provides a privacy-by-design infrastructure connecting FAIR (Findable, Accessible, Interoperable, Reusable) data sources and allows distributed data analysis and machine learning. Patient data never leaves a healthcare institute. MATERIALS AND METHODS: Lung cancer patient-specific databases (tumor staging and post-treatment survival information) of oncology departments were translated according to a FAIR data model and stored locally in a graph database. Software was installed locally to enable deployment of distributed machine learning algorithms via a central server. Algorithms (MATLAB, code and documentation publicly available) are patient privacy-preserving as only summary statistics and regression coefficients are exchanged with the central server. A logistic regression model to predict post-treatment two-year survival was trained and evaluated by receiver operating characteristic curves (ROC), root mean square prediction error (RMSE) and calibration plots. RESULTS: In 4 months, we connected databases with 23 203 patient cases across 8 healthcare institutes in 5 countries (Amsterdam, Cardiff, Maastricht, Manchester, Nijmegen, Rome, Rotterdam, Shanghai) using the PHT. Summary statistics were computed across databases. A distributed logistic regression model predicting post-treatment two-year survival was trained on 14 810 patients treated between 1978 and 2011 and validated on 8 393 patients treated between 2012 and 2015. CONCLUSION: The PHT infrastructure demonstrably overcomes patient privacy barriers to healthcare data sharing and enables fast data analyses across multiple institutes from different countries with different regulatory regimens. This infrastructure promotes global evidence-based medicine while prioritizing patient privacy.

Non-invasive imaging prediction of tumor hypoxia: A novel developed and externally validated CT and FDG-PET-based radiomic signatures.

BACKGROUND: Tumor hypoxia increases resistance to radiotherapy and systemic therapy. Our aim was to develop and validate a disease-agnostic and disease-specific CT (+FDG-PET) based radiomics hypoxia classification signature. MATERIAL AND METHODS: A total of 808 patients with imaging data were included: N = 100 training/N = 183 external validation cases for a disease-agnostic CT hypoxia classification signature, N = 76 training/N = 39 validation cases for the H&N CT signature and N = 62 training/N = 36 validation cases for the Lung CT signature. The primary gross tumor volumes (GTV) were manually defined by experts on CT. In order to dichotomize between hypoxic/well-oxygenated tumors a threshold of 20% was used for the [18F]-HX4-derived hypoxic fractions (HF). A random forest (RF)-based machine-learning classifier/regressor was trained to classify patients as hypoxia-positive/ negative based on radiomic features. RESULTS: A 11 feature "disease-agnostic CT model" reached AUC's of respectively 0.78 (95% confidence interval [CI], 0.62-0.94), 0.82 (95% CI, 0.67-0.96) and 0.78 (95% CI, 0.67-0.89) in three external validation datasets. A "disease-agnostic FDG-PET model" reached an AUC of 0.73 (0.95% CI, 0.49-0.97) in validation by combining 5 features. The highest "lung-specific CT model" reached an AUC of 0.80 (0.95% CI, 0.65-0.95) in validation with 4 CT features, while the "H&N-specific CT model" reached an AUC of 0.84 (0.95% CI, 0.64-1.00) in validation with 15 CT features. A tumor volume-alone model was unable to significantly classify patients as hypoxia-positive/ negative. A significant survival split (P = 0.037) was found between CT-classified hypoxia strata in an external H&N cohort (n = 517), while 117 significant hypoxia gene-CT signature feature associations were found in an external lung cohort (n = 80). CONCLUSION: The disease-specific radiomics signatures perform better than the disease agnostic ones. By identifying hypoxic patients our signatures have the potential to enrich interventional hypoxia-targeting trials.

Privacy-preserving distributed learning of radiomics to predict overall survival and HPV status in head and neck cancer.

A major challenge in radiomics is assembling data from multiple centers. Sharing data between hospitals is restricted by legal and ethical regulations. Distributed learning is a technique, enabling training models on multicenter data without data leaving the hospitals ("privacy-preserving" distributed learning). This study tested feasibility of distributed learning of radiomics data for prediction of two year overall survival and HPV status in head and neck cancer (HNC) patients. Pretreatment CT images were collected from 1174 HNC patients in 6 different cohorts. 981 radiomic features were extracted using Z-Rad software implementation. Hierarchical clustering was performed to preselect features. Classification was done using logistic regression. In the validation dataset, the receiver operating characteristics (ROC) were compared between the models trained in the centralized and distributed manner. No difference in ROC was observed with respect to feature selection. The logistic regression coefficients were identical between the methods (absolute difference <10-7). In comparison of the full workflow (feature selection and classification), no significant difference in ROC was found between centralized and distributed models for both studied endpoints (DeLong p > 0.05). In conclusion, both feature selection and classification are feasible in a distributed manner using radiomics data, which opens new possibility for training more reliable radiomics models.