Cancer Imaging Archive Research Articles

Abstract 6331: NCI’s Clinical Proteomic Tumor Analysis Consortium: A proteogenomic cancer analysis program

Abstract The National Cancer Institute’s Clinical Proteomic Tumor Analysis Consortium (NCI’s CPTAC) is an integrative proteogenomic program composed of a Proteogenomic Tumor Characterization Program and a Proteogenomic Translational Research Program (https://proteomics.cancer.gov). The goal of CPTAC is to improve prevention, early detection, diagnosis, and treatment of cancer by enhancing the understanding of the molecular mechanisms of cancer, advancing proteogenome science and technology development through community resources, and accelerating the translation of molecular findings into the clinic. Using state-of-the-art high throughput standardized mass spectrometry-based methods, the Proteogenomic Tumor Characterization Program performs deep comprehensive proteogenomic analysis of cancer types with all data and assays to be released to the research community. CPTAC teams have characterized a plethora of treatment-naïve tumor types including colorectal, ovarian, breast, clear cell renal cell carcinoma, uterine corpus endometrial carcinoma, lung adenocarcinoma, lung squamous cell carcinoma, brain (including adult, pediatric, and adolescent and young adult), head and neck squamous cell carcinoma, and pancreatic ductal adenocarcinoma. In the Proteogenomic Translational Research Program, CPTAC is partnering for the first time with NCI-sponsored clinical trials to support clinically relevant research projects that would elucidate biological mechanisms of therapeutic response, resistance, and/or toxicity. The Proteogenomic Translational Research Program currently explores triple negative breast cancer, high-grade serous ovarian cancer, and acute myeloid leukemia. All mass spectrometry-based proteomic, genomic, and imaging data (histopathology and radiology) are made publicly available at the CPTAC Proteomic Data Commons (https://pdc.cancer.gov), Genomic Data Commons (https://gdc.cancer.gov), and the Cancer Imaging Archive (https://cancerimagingarchive.net) respectively. CPTAC is also supporting development of new proteogenomic data analysis tools (https://proteomics.cancer.gov/resources/computational-tools). In addition, the CPTAC Assay Portal (http://assays.cancer.gov) is a public resource populated with mass spectrometry-based targeted proteomic assays developed by the consortium for quantitatively measuring proteins of interest, including those discovered through comprehensive tumor characterization. Lastly, well-characterized monoclonal antibodies targeting cancer-specific proteins and peptides are also made available at CPTAC’s Antibody Portal (http://antibodies.cancer.gov). Citation Format: Mehdi Mesri, Eunkyung An, Tara Hiltke, Ana I. Robles, Henry Rodriguez, CPTAC Investigators. NCI’s Clinical Proteomic Tumor Analysis Consortium: A proteogenomic cancer analysis program [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 6331.

Classification of acute myeloid leukemia M1 and M2 subtypes using machine learning

BackgroundClassification of acute myeloid leukemia (AML) relies on manual analysis of bone marrow or peripheral blood smear images. We aimed to construct a machine learning model for automatic classification of AML-M1 and M2 subtypes in bone marrow smear images. MethodsBone marrow smear images of AML patients were extracted from the Cancer Imaging Archive (TCIA) open database. Classification criteria of AML subtypes were based on the French-American-British (FAB) classification system. Random forest method and broad learning system (BLS) were used to develop the classification model. Morphological features, radiomics features, and clinical features were extracted. The performance of the classification model was evaluated by calculating accuracy, precision, recall, F1-score, and area under the curve (AUC). A total of 50 bone marrow smear images (AML-M1, 31 cases; AML-M2, 19 cases) with 500 slices were included in this study. ResultsA total of 43 morphological features, 276 radiomics features, and 1 clinical feature were extracted. Finally, 9 variables including 2 morphological features, 6 radiomics features, and 1 clinical feature were selected into the classification model. The best classification performance was observed in the random forest model with 9 variables, with the average accuracy, AUC, F1-score, recall, and precision of the model being 0.998 ± 0.003, 0.998 ± 0.004, 0.998 ± 0.004, 0.996 ± 0.009, and 1 ± 0, respectively. ConclusionThe random forest model performed well for the classification of AML-M1 and M2, which may provide a tool for clinicians to classify AML-M1 and M2.

Abstract 2459: Automated VIF methods improve DCE-MRI parameterization repeatability in GBM

Abstract Introduction: Analysis of Dynamic Contrast Enhanced (DCE) MRI is widely used for measuring perfusion kinetics of cancerous tissue, though its implementation and results may vary between sites. Here we study the repeatability of quantitative DCE-MRI parameterization in a publicly available Glioblastoma (GBM) double-baseline dataset and set forth percent repeatability coefficients (%RC) which constitute a level of change at which a detectable biological change in biology may be differentiated from experimental noise. Methods: Double-baseline imaging data was previously acquired for The Cancer Imaging Archive, in the Quantitative Imaging Network (QIN) GBM Treatment Response (TR) dataset. This data includes dual baseline acquisitions of DCE-MRI MRI data performed within 5 days, for 29 patients. Each acquisition is analyzed using two vascular input function (VIF) methods: a manually segmented VIF from the superior sagittal sinus, and an automatically derived VIF determined from the top 10% of rapidly enhancing voxels. Each dynamic scan is analyzed the the Extended Tofts-Kety model (eTM; Ktrans, ve, vp) pharmacokinetic (PK) model. Bland-Altman analysis is then performed on the log-transformed PK parameters to determine the percent repeatability coefficients (%RC) of the data (95% confidence), as well as bias between first and second DCE scans. This analysis was repeated for two separate observers, writing their own analysis code, both of which are validated on a digital reference object to ensure consistency between analysis methods with zero noise. Results: For bothVIF determination methods, the 25-75th percentiles of the change in PK parameters was less than +/- 25%, the standard set forth by QIBA as a detectable biological change. For most parameters, VIF methods, and model combinations, the automatic VIF method has smallest %RC, indicating that automatic VIF determination may enhance the repeatability of the measurements between timepoints. The %RC was lowest for vp when using an automatically determined VIF. Using the automatically determined VIF and eTM, the %RCs of Ktrans, ve, and vp were measured as 72%, 82%, and 53%, respectively. Using the eTM and manually segmented VIF, the %RCs of Ktrans, ve, and vp were measured as 83%, 88%, and 66%. Conclusions: Using the QIN-GBM-TR dataset, the authors have determined that the %RC of Ktrans, ve, and vp to be 72%, 82%, and 53%, respectively (eTM, auto-VIF). Because these relative changes are large, we recommend that all sites abide by QIBA guidelines to enhance DCE parameterization reproducibility. We also recommend usage of an automated AIF determination method, as opposed to manual segmentation of large blood vessels. Citation Format: Ryan T. Woodall, Prativa Sahoo, Yujie Cui, Bihong Chen, Mark Shiroishi, Cristina Lavini, Paul Frankel, Margarita Gutova, Christine Brown, Jennifer Munson, Russell Rockne. Automated VIF methods improve DCE-MRI parameterization repeatability in GBM [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 2459.

A multi-modal deep learning model integrates clinical, pathomic, and radiomic features for glioma classification and grading.

e14038 Background: Histopathologic glioma diagnosis relies on the interpretation of the morphologic appearance of tumor cells. This conventional approach suffers from documented inter-pathologist disagreements with minor discordances in 36% and serious discordances in 8.8% of cases (Bruner JM et al. Cancer. 1997). Machine learning applied to automatically extracted pathomic features offers the potential to reduce this inter-reader variability and improve classification accuracy. The addition of clinical history and radiomic features may further improve performance. This work compares the performance of automatic pathomic and radiomic feature analysis to human histologic impressions for glioma classification according to the 2021 WHO scheme. Methods: We developed and evaluated models for predicting the 2021 WHO Glioma class using data from The Cancer Genome Atlas and The Cancer Imaging Archive. We used a cohort of 112 glioma cases with complete data of molecular markers, digital slides, and MRI images. Using molecular markers, we divided this cohort into 2021 WHO types, including 61 glioblastomas, 38 astrocytomas, and 13 oligodendroglioma. We extracted 374 pathomic features summarizing the tumor-rich region on each digital slide. Neural network models were developed with pathomic features alone or in combination with clinical factors (sex, age, and Karnofsky index) and 670 radiomic features obtained from multi-sequence MRI exams. Accuracy was estimated with 5-fold cross validation. Feature importance analysis was carried out with recursive feature elimination to identify the top pathomic and radiomic features related to the specific glioma diagnosis. Results: In this cohort, the conventional histologic impression, reliant on pathologist interpretation, was successful in identifying the WHO glioma class with an accuracy of 0.77 +/- 0.08. The use of pathomics features alone, representing a fully automatic approach, had an accuracy of 0.83 +/- 0.08. The use of pathomics and radiomics information increased the accuracy to 0.88 +/- 0.08. Feature importance analysis revealed that the most important pathomic features were related to nuclear size, nuclear hematoxylin staining intensity, and nuclear circularity. Conclusions: In this pilot study, our machine learning model with pathomic features achieved promising performance for glioma classification. Combining clinical and radiomic features further enhanced performance. We found that the pathomic features that most correlated with the pathologic classification are the same key features that pathologists use in their diagnostic evaluation. While molecular testing is now the definitive reference information for final glioma classification, this pilot study suggests that pathomics analysis could have a valuable role during a pathologist’s initial histologic review.

AI in NSCLC: PET-CT & histology model.

e21044 Background: Almost 70% of cases of lung cancer are diagnosed at advanced stage, albeit with sophisticated imaging modalities available. Of these, 35% are tiny nodules that are often missed at initial radiological screening, owing to limitations of resolution of human vision.Dramatic shift in therapeutic paradigm in the management of lung cancer in the last decade has ushered an era of stratified medicine. In India, one-third population inhabits rural lands;lack of awareness for molecular medicine poses challenges to treating oncologists. AI-based cancer detection and prediction of potential oncogenic driver at diagnosis, will help alleviate patient anxiety caused due to diagnostic delays, aiding in prompt initiation of optimal therapy.Here, a comprehensive deep-CNN based diagnosis of lung adenocarcinoma (LUAD) was implemented using both PET/CT and histopathology images of 100 patients. Methods: Data collection: 143 whole-slide images were downloaded from Dartmouth Lung Cancer Histology Dataset from Biomedical Informatics Research and Data Science website and PET/CT dataset-A Large-Scale CT and PET/CT Dataset for Lung Cancer Diagnosis (Lung-PET-CT-Dx) from The Cancer Imaging Archive. This comprises 251,135 images from 355 participants across 436 studies. Creation of Deep-CNN model: Residual Neural Network - 18 was used for histopathlogic classification followed by Detectron2’s region-based deep-CNN for detection of presence of tumours in lung PET/CT images. Properties such as size, morphology, attenuation, malignancy, and spiculation were also measured. The 18-layer deep network, ResNet-18, detected the histologic patterns using whole-slide histopathology images. Results: Results for 100 patients whose PET/CT images, histopathology images, and mutational status of three genes (EGFR, ALK, ROS) were readily available. The histologic patterns of tumour cells appear to be related to the mutational status of genes. The ResNet-18 CNN could predict various histologic patterns (Acinar, Lepidic, Papillary, Micropapillary, and Solid patterns). Conclusions: CNN models can help in early diagnosis and prompt treatment of lung cancer, if fairly accurate. Larger datasets are needed to improve the same. The final validation of the model is underway and will be reported in near future.[Table: see text]

Computerized Classification Method for 1p/19q Codeletion in Low-Grade Glioma on Brain MRI Using Multi-Scale 3D-CNNs with Attention Mechanism

The purpose of this study was to develop a computerized classification method for low-grade gliomas (LGGs) with/without 1p/19q codeletion on brain MRI using multi-scale 3D-convolutional neural networks (Multi-scale 3D-CNNs) with an attention mechanism that analyzes only the tumor region. Our database consisted of brain T2-weighted MRI images for 159 patients (102 LGGs with 1p/19q codeletion and 57 LGGs without it) from The Cancer Imaging Archive. The proposed method was constructed from a feature extractor, an attention mechanism, and a perception branch. In the feature extractor, the feature maps were extracted from input images. The attention mechanism generated the attention maps focusing on a tumor region from those feature maps. The feature maps on the tumor region were then obtained using an attention pooling with the attention maps. In the perception branch, the likelihood of LGG with 1p/19q codeletion was evaluated based on the feature maps of the tumor region. The classification accuracy, the sensitivity, the specificity, and the area under the receiver operating characteristic curve for the proposed method were 78.0%, 82.4%, 70.2%, and 0.838, showing a significant improvement when compared with the multi-scale 3D-CNNs without the attention mechanism (69.8%, 75.5%, 59.6%, and 0.717; p = 0.001).

Hepatic Arterial Patterns and Variations: Using CT Data Visualization Software to Decrease Procedural Complications Due to Hepatic Arterial Variations

IntroThree‐dimensional (3D) models have recently been used in pre‐operative planning in rare, complex surgeries in many disciplines. The goal of this study was to determine the efficacy of using 3D models in common procedures that involve increased risk of injury to hepatic arteries in the biliary region. This would allow surgeons to plan around individual variation and prevent intra‐operative complications in these regions. This study focuses on the feasibility of converting computed tomography (CT) scans into 3D models to identify anatomical variation and compares the results to the published rates of anatomical variation for the structures considered.MethodsA total of 27 (n=27) CT scans of healthy individuals were obtained from the Cancer Imaging Archive and digitally segmented with Dragonfly and the 3D Slicer software using the semi‐automated grow from seeds tool. The results were then manually reviewed and edited for accuracy. The liver, gall bladder, and spleen were segmented as well as the associated hepatic arterial vasculature: celiac trunk, splenic artery, left gastric artery, common hepatic artery, gastroduodenal artery, proper hepatic artery, right and left hepatic arteries, and the superior mesenteric artery. The segmented scans were then converted to a 3D model where they were classified using the CRL classification system which has a 99.8% classification rate.ResultsAll 3D models were classified into a CRL classification type. Nineteen (n=19) models were classified as Type 1, two (n=2) models as Type 2, one (n=1) model as Type 3, one (n=1) model as Type 4, two (n=2) models as Type 5, one (n=1) model as Type 9, and one (n=1) model did not fit under the classification system and remained unclassified. Previously published data was used to calculate the statistical expected value of each CRL type. A chi‐square analysis was performed to determine if the dataset fit normal patterns of variation. The results of the statistical test were statistically different χ2=0.0008 than previously published patterns of anatomical variation due to the rare unclassifiable variant.ConclusionThe study showed that it is feasible to use pre‐operative CT scans to reconstruct 3D models that can be used to identify variant anatomy pre‐operatively to assist in the planning of complex surgeries. The 3D slicer software did not require a computer with substantial hardware requirements and is able to be used on a personal computer. This means it is not necessary to purchase special equipment to implement in the hospital and could potentially allow physicians to explain procedures to patients in clinics outside of hospitals. In order to expand on the capabilities of pre‐operative 3D models, it is necessary to investigate more automated methods to accurately segment the appropriate anatomy.

A Machine Learning Model Based on PET/CT Radiomics and Clinical Characteristics Predicts Tumor Immune Profiles in Non-Small Cell Lung Cancer: A Retrospective Multicohort Study.

BackgroundThe tumor immune microenvironment (TIME) phenotypes have been reported to mainly impact the efficacy of immunotherapy. Given the increasing use of immunotherapy in cancers, knowing an individual’s TIME phenotypes could be helpful in screening patients who are more likely to respond to immunotherapy. Our study intended to establish, validate, and apply a machine learning model to predict TIME profiles in non-small cell lung cancer (NSCLC) by using 18F-FDG PET/CT radiomics and clinical characteristics.MethodsThe RNA-seq data of 1145 NSCLC patients from The Cancer Genome Atlas (TCGA) cohort were analyzed. Then, 221 NSCLC patients from Daping Hospital (DPH) cohort received18F-FDG PET/CT scans before treatment and CD8 expression of the tumor samples were tested. The Artificial Intelligence Kit software was used to extract radiomic features of PET/CT images and develop a radiomics signature. The models were established by radiomics, clinical features, and radiomics-clinical combination, respectively, the performance of which was calculated by receiver operating curves (ROCs) and compared by DeLong test. Moreover, based on radiomics score (Rad-score) and clinical features, a nomogram was established. Finally, we applied the combined model to evaluate TIME phenotypes of NSCLC patients in The Cancer Imaging Archive (TCIA) cohort (n = 39).ResultsTCGA data showed CD8 expression could represent the TIME profiles in NSCLC. In DPH cohort, PET/CT radiomics model outperformed CT model (AUC: 0.907 vs. 0.861, P = 0.0314) to predict CD8 expression. Further, PET/CT radiomics-clinical combined model (AUC = 0.932) outperformed PET/CT radiomics model (AUC = 0.907, P = 0.0326) or clinical model (AUC = 0.868, P = 0.0036) to predict CD8 expression. In the TCIA cohort, the predicted CD8-high group had significantly higher immune scores and more activated immune pathways than the predicted CD8-low group (P = 0.0421).ConclusionOur study indicates that 18F-FDG PET/CT radiomics-clinical combined model could be a clinically practical method to non-invasively detect the tumor immune status in NSCLCs.

Limited parameter denoising for low-dose X-ray computed tomography using deep reinforcement learning.

The use of deep learning has successfully solved several problems in the field of medical imaging. Deep learning has been applied to the CT denoising problem successfully. However, the use of deep learning requires large amounts of data to train deep convolutional networks (CNNs). Moreover, due to the large parameter count, such deep CNNs may cause unexpected results. In this study, we introduce a novel CT denoising framework, which has interpretable behavior and provides useful results with limited data. We employ bilateral filtering in both the projection and volume domains to remove noise. To account for nonstationary noise, we tune the σ parameters of the volume for every projection view and every volume pixel. The tuning is carried out by two deep CNNs. Due to the impracticality of labeling, the two-deep CNNs are trained via a Deep-Q reinforcement learning task. The reward for the task is generated by using a custom reward function represented by a neural network. Our experiments were carried out on abdominal scans for the Mayo Clinic the cancer imaging archive (TCIA) dataset and the American association of physicists in medicine (AAPM) Low Dose CT Grand Challenge. Our denoising framework has excellent denoising performance increasing the peak signal to noise ratio (PSNR) from 28.53 to 28.93 and increasing the structural similarity index (SSIM) from 0.8952 to 0.9204. We outperform several state-of-the-art deep CNNs, which have several orders of magnitude higher number of parameters (p-value [PSNR] = 0.000, p-value [SSIM] = 0.000). Our method does not introduce any blurring, which is introduced by mean squared error (MSE) loss-based methods, or any deep learning artifacts, which are introduced by wasserstein generative adversarial network (WGAN)-based models. Our ablation studies show that parameter tuning and using our reward network results in the best possible results. We present a novel CT denoising framework, which focuses on interpretability to deliver good denoising performance, especially with limited data. Our method outperforms state-of-the-art deep neural networks. Future work will be focused on accelerating our method and generalizing it to different geometries and bodyparts.

External Validation of a Convolutional Neural Network for IDH Mutation Prediction.

Background and Objectives: The IDH (isocitrate dehydrogenase) status represents one of the main prognosis factors for gliomas. However, determining it requires invasive procedures and specialized surgical skills. Medical imaging such as MRI is essential in glioma diagnosis and management. Lately, fields such as Radiomics and Radiogenomics emerged as pertinent prediction tools for extracting molecular information out of medical images. These fields are based on Artificial Intelligence algorithms that require external validation in order to evaluate their general performance. The aim of this study was to provide an external validation for the algorithm formulated by Yoon Choi et al. of IDH status prediction using preoperative common MRI sequences and patient age. Material and Methods: We applied Choi’s IDH status prediction algorithm on T1c, T2 and FLAIR preoperative MRI images of gliomas (grades WHO II-IV) of 21 operated adult patients from the Neurosurgery clinic of the Cluj County Emergency Clinical Hospital (CCECH), Cluj-Napoca Romania. We created a script to automate the testing process with DICOM format MRI sequences as input and IDH predicted status as output. Results: In terms of patient characteristics, the mean age was 48.6 ± 15.6; 57% were female and 43% male; 43% were IDH positive and 57% IDH negative. The proportions of WHO grades were 24%, 14% and 62% for II, III and IV, respectively. The validation test achieved a relative accuracy of 76% with 95% CI of (53%, 92%) and an Area Under the Curve (AUC) through DeLong et al. method of 0.74 with 95% CI of (0.53, 0.91) and a p of 0.021. Sensitivity and Specificity were 0.78 with 95% CI of (0.45, 0.96) and 0.75 with 95% CI of (0.47, 0.91), respectively. Conclusions: Although our results match the external test the author made on The Cancer Imaging Archive (TCIA) online dataset, performance of the algorithm on external data is still not high enough for clinical application. Radiogenomic approaches remain a high interest research field that may provide a rapid and accurate diagnosis and prognosis of patients with intracranial glioma.

Effects of Anatomical Variations of the Stomach on Body-Surface Gastric Mapping Investigated Using a Large Population-Based Multiscale Simulation Approach.

Body-surface gastric mapping (BSGM) measures the resultant body-surface potentials of gastric slow waves using an array of cutaneous electrodes. However, there is no established protocol to guide the placement of the mapping array and to account for the effects of biodiversity on the interpretation of BSGM data. This study aims to quantify the effect of anatomical variation of the stomach on body surface potentials. To this end, 93 subject specific models of the stomach and torso were developed, based on data obtained from the Cancer Imaging Archive. For each subject a set of points were created to model general anatomy the stomach and the torso, using a finite element mesh. A bidomain model was used to simulate the gastric slow waves in the antegrade wave (AW) direction and formation of colliding waves (CW). A forward modeling approach was employed to simulate body-surface potentials from the equivaelent dipoles. Simulated data were sampled from a 5 × 5 array of electrodes from the body-surface and compared between AW and CW cases. Anatomical parameters such as the Euclidean distance from the xiphoid process (8.6 ± 2.2 cm), orientation relative to the axial plane (195 ± 20.0°) were quantified. Electrophysiological simulations of AW and CW were both correlated to specific metrics derived from BSGM signals. In general, the maximum amplitude ( ∆ ϕ) and orientation ( θ) of the signals provided consistent separation of AW and CW. The findings of this study will aid gastric BSGM electrode array design and placement protocol in clinical practices.

C-NMC: B-lineage acute lymphoblastic leukaemia: A blood cancer dataset

Development of computer-aided cancer diagnostic tools is an active research area owing to the advancements in deep-learning domain. Such technological solutions provide affordable and easily deployable diagnostic tools. Leukaemia, or blood cancer, is one of the leading cancers causing more than 0.3 million deaths every year. In order to aid the development of such an AI-enabled tool, we collected and curated a microscopic image dataset, namely C-NMC, of more than 15000 cancer cell images at a very high resolution of B-Lineage Acute Lymphoblastic Leukaemia (B-ALL). The dataset is prepared at the subject-level and contains images of both healthy and cancer patients. So far, this is the largest (as well as curated) dataset on B-ALL cancer in the public domain. C-NMC is available at The Cancer Imaging Archive (TCIA), USA and can be helpful for the research community worldwide for the development of B-ALL cancer diagnostic tools. This dataset was utilized in an international medical imaging challenge held at ISBI 2019 conference in Venice, Italy. In this paper, we present a detailed description and challenges of this dataset. We also present benchmarking results of all the methods applied so far on this dataset.

Context-Aware Saliency Guided Radiomics: Application to Prediction of Outcome and HPV-Status from Multi-Center PET/CT Images of Head and Neck Cancer.

Simple SummaryThis study investigated the ability of context-aware saliency-guided PET/CT radiomics in the prediction of outcome and HPV status for head and neck cancer. In total, 806 HNC patients (training vs. validation vs. external testing: 500 vs. 97 vs. 209) from 9 centers were collected from The Cancer Imaging Archive (TCIA). Saliency-guided radiomics showed enhanced performance for both outcome and HPV-status predictions relative to conventional radiomics. The radiomics-predicted HPV status also showed complementary prognostic value. This multi-center study highlights the feasibility of saliency-guided PET/CT radiomics in outcome predictions of head and neck cancer, confirming that certain regions are more relevant to tumor aggressiveness and prognosis.Purpose: This multi-center study aims to investigate the prognostic value of context-aware saliency-guided radiomics in 18F-FDG PET/CT images of head and neck cancer (HNC). Methods: 806 HNC patients (training vs. validation vs. external testing: 500 vs. 97 vs. 209) from 9 centers were collected from The Cancer Imaging Archive (TCIA). There were 100/384 and 60/123 oropharyngeal carcinoma (OPC) patients with human papillomavirus (HPV) status in training and testing cohorts, respectively. Six types of images were used for radiomics feature extraction and further model construction, namely (i) the original image (Origin), (ii) a context-aware saliency map (SalMap), (iii, iv) high- or low-saliency regions in the original image (highSal or lowSal), (v) a saliency-weighted image (SalxImg), and finally, (vi) a fused PET-CT image (FusedImg). Four outcomes were evaluated, i.e., recurrence-free survival (RFS), metastasis-free survival (MFS), overall survival (OS), and disease-free survival (DFS), respectively. Multivariate Cox analysis and logistic regression were adopted to construct radiomics scores for the prediction of outcome (Rad_Ocm) and HPV-status (Rad_HPV), respectively. Besides, the prognostic value of their integration (Rad_Ocm_HPV) was also investigated. Results: In the external testing cohort, compared with the Origin model, SalMap and SalxImg achieved the highest C-indices for RFS (0.621 vs. 0.559) and MFS (0.785 vs. 0.739) predictions, respectively, while FusedImg performed the best for both OS (0.685 vs. 0.659) and DFS (0.641 vs. 0.582) predictions. In the OPC HPV testing cohort, FusedImg showed higher AUC for HPV-status prediction compared with the Origin model (0.653 vs. 0.484). In the OPC testing cohort, compared with Rad_Ocm or Rad_HPV alone, Rad_Ocm_HPV performed the best for OS and DFS predictions with C-indices of 0.702 (p = 0.002) and 0.684 (p = 0.006), respectively. Conclusion: Saliency-guided radiomics showed enhanced performance for both outcome and HPV-status predictions relative to conventional radiomics. The radiomics-predicted HPV status also showed complementary prognostic value.

Mining whole-lung information by artificial intelligence for predicting EGFR genotype and targeted therapy response in lung cancer: a multicohort study

Epidermal growth factor receptor (EGFR) genotype is crucial for treatment decision making in lung cancer, but it can be affected by tumour heterogeneity and invasive biopsy during gene sequencing. Importantly, not all patients with an EGFR mutation have good prognosis with EGFR-tyrosine kinase inhibitors (TKIs), indicating the necessity of stratifying for EGFR-mutant genotype. In this study, we proposed a fully automated artificial intelligence system (FAIS) that mines whole-lung information from CT images to predict EGFR genotype and prognosis with EGFR-TKI treatment. We included 18 232 patients with lung cancer with CT imaging and EGFR gene sequencing from nine cohorts in China and the USA, including a prospective cohort in an Asian population (n=891) and The Cancer Imaging Archive cohort in a White population. These cohorts were divided into thick CT group and thin CT group. The FAIS was built for predicting EGFR genotype and progression-free survival of patients receiving EGFR-TKIs, and it was evaluated by area under the curve (AUC) and Kaplan-Meier analysis. We further built two tumour-based deep learning models as comparison with the FAIS, and we explored the value of combining FAIS and clinical factors (the FAIS-C model). Additionally, we included 891 patients with 56-panel next-generation sequencing and 87 patients with RNA sequencing data to explore the biological mechanisms of FAIS. FAIS achieved AUCs ranging from 0·748 to 0·813 in the six retrospective and prospective testing cohorts, outperforming the commonly used tumour-based deep learning model. Genotype predicted by the FAIS-C model was significantly associated with prognosis to EGFR-TKIs treatment (log-rank p<0·05), an important complement to gene sequencing. Moreover, we found 29 prognostic deep learning features in FAIS that were able to identify patients with an EGFR mutation at high risk of TKI resistance. These features showed strong associations with multiple genotypes (p<0·05, t test or Wilcoxon test) and gene pathways linked to drug resistance and cancer progression mechanisms. FAIS provides a non-invasive method to detect EGFR genotype and identify patients with an EGFR mutation at high risk of TKI resistance. The superior performance of FAIS over tumour-based deep learning methods suggests that genotype and prognostic information could be obtained from the whole lung instead of only tumour tissues. National Natural Science Foundation of China.

Radiomics predicts the prognosis of patients with locally advanced breast cancer by reflecting the heterogeneity of tumor cells and the tumor microenvironment

BackgroundThis study investigated the efficacy of radiomics to predict survival outcome for locally advanced breast cancer (LABC) patients and the association of radiomics with tumor heterogeneity and microenvironment.MethodsPatients with LABC from 2010 to 2015 were retrospectively reviewed. Radiomics features were extracted from enhanced MRI. We constructed the radiomics score using lasso and assessed its prognostic value. An external validation cohort from The Cancer Imaging Archive was used to assess phenotype reproducibility. Sequencing data from TCGA and our center were applied to reveal genomic landscape of different radiomics score groups. Tumor infiltrating lymphocytes map and bioinformatics methods were applied to evaluate the heterogeneity of tumor microenvironment. Computational histopathology was also applied.ResultsA total of 278 patients were divided into training cohort and validation cohort. Radiomics score was constructed and significantly associated with disease-free survival (DFS) of the patients in training cohort, validation cohort and external validation cohort (p < 0.001, p = 0.014 and p = 0.041, respectively). The radiomics-based nomogram showed better predictive performance of DFS compared with TNM model. Distinct gene expression patterns were identified. Immunophenotype and immune cell composition was different in each radiomics score group. The link between radiomics and computational histopathology was revealed.ConclusionsThe radiomics score could effectively predict prognosis of LABC after neoadjuvant chemotherapy and radiotherapy. Radiomics revealed heterogeneity of tumor cell and tumor microenvironment and holds great potential to facilitate individualized DFS estimation and guide personalized care.

A deep learning pipeline to simulate fluorodeoxyglucose (FDG) uptake in head and neck cancers using non-contrast CT images without the administration of radioactive tracer

ObjectivesPositron emission tomography (PET) imaging is a costly tracer-based imaging modality used to visualise abnormal metabolic activity for the management of malignancies. The objective of this study is to demonstrate that non-contrast CTs alone can be used to differentiate regions with different Fluorodeoxyglucose (FDG) uptake and simulate PET images to guide clinical management.MethodsPaired FDG-PET and CT images (n = 298 patients) with diagnosed head and neck squamous cell carcinoma (HNSCC) were obtained from The cancer imaging archive. Random forest (RF) classification of CT-derived radiomic features was used to differentiate metabolically active (tumour) and inactive tissues (ex. thyroid tissue). Subsequently, a deep learning generative adversarial network (GAN) was trained for this CT to PET transformation task without tracer injection. The simulated PET images were evaluated for technical accuracy (PERCIST v.1 criteria) and their ability to predict clinical outcome [(1) locoregional recurrence, (2) distant metastasis and (3) patient survival].ResultsFrom 298 patients, 683 hot spots of elevated FDG uptake (elevated SUV, 6.03 ± 1.71) were identified. RF models of intensity-based CT-derived radiomic features were able to differentiate regions of negligible, low and elevated FDG uptake within and surrounding the tumour. Using the GAN-simulated PET image alone, we were able to predict clinical outcome to the same accuracy as that achieved using FDG-PET images.ConclusionThis pipeline demonstrates a deep learning methodology to simulate PET images from CT images in HNSCC without the use of radioactive tracer. The same pipeline can be applied to other pathologies that require PET imaging.

Efficient Radiomics-Based Classification of Multi-Parametric MR Images to Identify Volumetric Habitats and Signatures in Glioblastoma: A Machine Learning Approach.

Simple SummaryGlioblastomas carry a poor prognosis and usually presents with heterogeneous regions in the brain tumor. Multi-parametric MR images can show morphological characteristics. Radiomics features refer to the extraction of a large number of quantitative measurements that describe the geometry, intensity, and texture which were extracted from contrast-enhanced T1-weighted images from anatomical MRI and metabolic features from PET. It also provides a qualitative image interpretation as well as cellular, molecular, and tumor properties. Thus, it derives additional information about the entire tumor volume which is generally of irregular shape and size from routinely evaluated “non-invasive” imaging biomarkers techniques. We demonstrated volumetric habitats and signatures in necrosis, solid tumor, peritumoral tissue, and edema with key biological processes and phenotype features. This provides physicians with key information on how the disease is progressing in the brain and can also give an indication of how well treatment is working.Glioblastoma (GBM) is a fast-growing and aggressive brain tumor of the central nervous system. It encroaches on brain tissue with heterogeneous regions of a necrotic core, solid part, peritumoral tissue, and edema. This study provided qualitative image interpretation in GBM subregions and radiomics features in quantitative usage of image analysis, as well as ratios of these tumor components. The aim of this study was to assess the potential of multi-parametric MR fingerprinting with volumetric tumor phenotype and radiomic features to underlie biological process and prognostic status of patients with cerebral gliomas. Based on efficiently classified and retrieved cerebral multi-parametric MRI, all data were analyzed to derive volume-based data of the entire tumor from local cohorts and The Cancer Imaging Archive (TCIA) cohorts with GBM. Edema was mainly enriched for homeostasis whereas necrosis was associated with texture features. The proportional volume size of the edema was about 1.5 times larger than the size of the solid part tumor. The volume size of the solid part was approximately 0.7 times in the necrosis area. Therefore, the multi-parametric MRI-based radiomics model reveals efficiently classified tumor subregions of GBM and suggests that prognostic radiomic features from routine MRI examination may also be significantly associated with key biological processes as a practical imaging biomarker.

Association testing for binary trees-A Markov branching process approach.

We propose a new approach to test associations between binary trees and covariates. In this approach, binary‐tree structured data are treated as sample paths of binary fission Markov branching processes (bMBP). We propose a generalized linear regression model and developed inference procedures for association testing, including variable selection and estimation of covariate effects. Simulation studies show that these procedures are able to accurately identify covariates that are associated with the binary tree structure by impacting the rate parameter of the bMBP. The problem of association testing on binary trees is motivated by modeling hierarchical clustering dendrograms of pixel intensities in biomedical images. By using semi‐synthetic data generated from a real brain‐tumor image, our simulation studies show that the bMBP model is able to capture the characteristics of dendrogram trees in brain‐tumor images. Our final analysis of the glioblastoma multiforme brain‐tumor data from The Cancer Imaging Archive identified multiple clinical and genetic variables that are potentially associated with brain‐tumor heterogeneity.

Automated pancreas segmentation and volumetry using deep neural network on computed tomography

Pancreas segmentation is necessary for observing lesions, analyzing anatomical structures, and predicting patient prognosis. Therefore, various studies have designed segmentation models based on convolutional neural networks for pancreas segmentation. However, the deep learning approach is limited by a lack of data, and studies conducted on a large computed tomography dataset are scarce. Therefore, this study aims to perform deep-learning-based semantic segmentation on 1006 participants and evaluate the automatic segmentation performance of the pancreas via four individual three-dimensional segmentation networks. In this study, we performed internal validation with 1,006 patients and external validation using the cancer imaging archive pancreas dataset. We obtained mean precision, recall, and dice similarity coefficients of 0.869, 0.842, and 0.842, respectively, for internal validation via a relevant approach among the four deep learning networks. Using the external dataset, the deep learning network achieved mean precision, recall, and dice similarity coefficients of 0.779, 0.749, and 0.735, respectively. We expect that generalized deep-learning-based systems can assist clinical decisions by providing accurate pancreatic segmentation and quantitative information of the pancreas for abdominal computed tomography.

Development of a radiomic–clinical nomogram for prediction of survival in patients with serous ovarian cancer

To develop and validate a radiomic-clinical nomogram to evaluate overall survival (OS) postoperatively in patients with serous ovarian cancer. Eighty serous ovarian cancer patients from The Cancer Imaging Archive (TCIA) database were used as the training set, and 39 eligible patients treated at Affiliated Huadu Hospital were used as the independent validation set. In total, 1,301 radiomics features were extracted from ovarian cancer lesions on venous-phase computed tomography (CT) images. Then, a radiomics signature was developed using the least absolute shrinkage and selection operator (LASSO) Cox regression algorithm in the training set. Moreover, a radiomic-clinical nomogram was constructed incorporating the radiomics signature and clinical predictors based on a multivariable Cox regression analysis. The performance of the nomogram was evaluated. Consisting of three selected features, the radiomics signature showed good discrimination in the training and validation sets with C-indexes of 0.694 (95% confidence interval [CI]: 0.613-0.775) and 0.709 (95% CI: 0.517-0.901), respectively. The radiomic-clinical nomogram contained the radiomics signature and four clinical predictors, including age, tumour size, pathological staging, and tumour grade. The nomogram showed favourable discrimination in the training set (C-index [95% CI], 0.754 [0.678-0.830]), which was confirmed in the validation set (C-index [95% CI], 0.727 [0.569-0.885]). According to the model, all patients were classified into high-risk and low-risk groups. Kaplan-Meier curves showed that there was a significant distinction between the OS of the high-risk and low-risk patients. The proposed radiomic-clinical nomogram can increase the predictive accuracy of OS in patients with serous ovarian cancer after surgery, which may aid in clinical decision-making.