Metastatic non-small cell lung cancer (NSCLC) and brain edema: a topographical and clinicopathological investigation utilizing deep learning-based artificial intelligence (DLBAI)

Metastatic non-small cell lung cancer (NSCLC) and brain edema: a topographical and clinicopathological investigation utilizing deep learning-based artificial intelligence (DLBAI) | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Metastatic non-small cell lung cancer (NSCLC) and brain edema: a topographical and clinicopathological investigation utilizing deep learning-based artificial intelligence (DLBAI) Jonathan Yun, Kristina Kurker, Georgios Maragkos, Jeyan Kumar, and 9 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3851661/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Tumor-associated vasogenic brain edema is a well-known contributor of morbidity and mortality in patients with metastatic disease to the brain. It is widely accepted that brain metastases (BM) is associated with extensive edema and can cause increased symptomatology such as pain, neurologic deficit, and elevated intracranial pressure depending on extent and location. We present a proof-of-concept retrospective analysis utilizing DLBAI to segment and detect radiological and topographical patterns of peritumoral edema and assess for clinicopathological correlates in 84 patients with NSCLC and BM who underwent surgical resection and were not previously on steroids. We found that overall, tumors in all locations demonstrated a mean 10:1 edema to tumor ratio (ETR) and an occipital tumor location was associated with a significantly elevated ETR. Within our cohort there were no other factors that were significantly associated with ETR. This study demonstrates a proof-of-concept that DLBAI is an efficient and accurate method of radiographic analysis that can be applied to detect and potentially predict clinicopathological data and prognostic determinants. Clinically, we demonstrate that NSCLC is associated with significant peritumoral edema and that topographical factors may be associated with increased extent of edema. Biological sciences/Computational biology and bioinformatics Biological sciences/Neuroscience Health sciences/Medical research Health sciences/Neurology Health sciences/Oncology Figures Figure 1 Introduction Brain metastases (BMs) are the most common adult intracranial malignancy. BMs will occur in approximately 10–20% of all adult cancer patients and portend a poor prognosis with an estimated median survival of 5 months for synchronous BMs. The epidemiological data is however limited because, unlike most cancers, there is no systematic nationwide reporting of secondary brain malignancies [ 1 – 3 ]. The lung is the most common site of origin for BMs, which are found in up to 20% of patients at the time of initial cancer diagnosis [ 2 ]. Non-small-cell lung cancer (NSCLC) is the most common form of lung cancer and up to 40% of patients with NSCLC will develop BMs during their disease course [ 4 ]. It is well established that BMs are associated with extensive peritumoral edema, which is a significant contributor to their morbidity and mortality. Symptomatology such as neurological deficit and pain secondary to increased intracranial pressure or mass effect can greatly affect both quality of life and overall survival. Pathogenesis of peritumoral edema in BM remains yet to be fully elucidated but is postulated to be secondary to tumor-induced angiogenesis and subsequent alterations in the surrounding microenvironment [ 5 – 7 ]. Understanding both radiographic and clinicopathologic patterns of peritumoral edema of BMs may further elucidate its underlying pathogenesis. To date, primary radiographic assessment of brain tumors is qualitative in nature, using freehand unidimensional and bi-dimensional measurements of specific areas of interest, which can be susceptible to inter-observer and intra-observer variability. Deep learning-based artificial intelligence (DLBAI) tumor segmentation algorithms have recently gained attention to circumvent the time-consuming and potentially biased manual assessment [ 8 – 10 ]. In this study, we investigate a series of patients with NSCLC with the use of AI-assisted brain tumor segmentation to detect significant topographic patterns of peritumoral edema of BMs, thus demonstrating a proof-of-concept clinical application of these deep learning algorithms. Results Baseline Characteristics A total of 84 patients with a surgically resected brain metastases were included for analysis. There were 46 female (54.1%) and 38 male (45.9%) patients. Mean age was 66.2 ± 9.6 years. Table 1 demonstrates tumor location characteristics for the patient cohort. Out of the resected lesions, 32 (38.1%) were in the frontal lobe, 8 (9.5%) in the temporal, 9 (10.7%) in the parietal, 8 (9.5%) in the occipital lobe, and 27 (32.1%) were in the cerebellum. Lesions located in the insula were classified as temporal and lesions located in the basal ganglia and thalamus were classified as frontal. There were no brain stem lesions identified in our cohort. Of the 84 patients, 21 (25%) had left-sided lesions, 36 (42.9%) right-sided, and 27 (32.1%) midline lesions. A total of 64 patients (81%) had at least one supratentorial lesion, and 37 (46.8%) had at least one infratentorial lesion. Thirty-six patients (45.6%) had multiple lesions. Eight patients were undergoing systemic therapy at the time of cranial diagnosis. Four patients were on pembrolizumab, 3 patients were on nivolumab, and 1 patient was on atezolizumab. There were no patients on anti-VEGF therapy at the time of diagnosis of their intracranial disease. Table 1 Patient characteristics. For each patient, age, sex, comorbidities, molecular markers of their cancer, and pre-diagnosis symptoms were recorded and compared. Variable Patients, n (%) Total 84 Female 46 (54.1) Comorbidities Hypertension 38 (44.7) Diabetes 10 (11.8) Cardiac 12 (14.1) Asthma/pulmonary 38 (44.7) Smoking 82 (96.5) Molecular markers EGFR mutant 4 (4.7) KRAS mutant 9 (10.6) PD-L1 expression > 50% 12 (14.1) ROS-1 mutant 1 (1.1) ALK mutant 0 (0) Symptoms Headaches 37 (44.1) Nausea 24 (28.6) Memory 15 (17.9) Confusion 25 (29.8) Focal deficit 38 (45.2) Balance issues 38 (45.2) Seizures 15 (17.9) Cranial nerve palsy 2 (2.4) Mean ± SD Age (in years) 66.2 ± 9.6 Regarding patient-specific comorbidities, 38 patients (44.7%) had hypertension, 10 patients (11.8%) had diabetes, 12 patients (14.1%) had cardiac comorbidities, 38 patients (44.7%) had asthma or other pulmonary past medical history, and 82 patients (96.5%) were past or present smokers. Thirty-seven patients (44.1%) had headaches, 24 patients (28.6%) complained of nausea, 15 patients (17.9%) had memory issues, 25 patients (29.8%) had confusion, 38 patients (45.2%) had a focal deficit, another 38 patients (45.2%) had balance issues, 15 patients (17.9%) had seizures, and 2 patients (2.4%) had a cranial nerve palsy. Table 2 presents patient-specific baseline characteristics. Table 2 Tumor location characteristics and relationships with tumoral edema. The total number of patients and percent of patients was calculated for which lobes the metastatic lesions were located, sidedness of their location, whether they were supratentorial or infratentorial, and whether the patient had multiple metastases. Univariate linear regression analysis was conducted to detect associations between each location variable and tumor edema. The lobe the tumors were in, sidedness of the location, compartment, and whether there were multiple metastases were evaluated for potential associations with tumor edema. P-values in bold are statistically significant. Variable Patients n (%) Regression Coefficient to Edema Volume 95% Confidence Interval P-value Location Frontal 32 (38.1) 21.61 -3.8 to 47 0.095 Temporal 8 (9.5) 29.9 -12.4 to 72.2 0.16 Parietal 9 (10.7) 10.38 -30.2 to 50.9 0.61 Occipital 8 (9.5) 34.34 -7.8 to 76.5 0.11 Cerebellar 27 (32.1) -53.3 -77.5 to -29.1 < 0.005 Side Left 21 (25) 29.06 0.8 to 57.4 0.044 Right 36 (42.9) 25.22 0.4 to 50 0.046 Midline 27 (32.1) -53.3 -77.5 to -29.1 < 0.005 Compartment Supratentorial 64 (81) 46.6 16.2 to 77 0.003 Infratentorial 37 (46.8) -24.63 -49.3 to 0 0.05 Multiple 36 (45.6) 17.34 -7.7 to 42.4 0.17 Regarding tumor histology, 61 patients (73%) had adenocarcinoma, 12 patients (14%) had squamous cell carcinoma, 7 patients (8%) had poorly differentiated NSCLC, 2 patients (2%) had adenosquamous carcinoma, and 2 patients (2%) had pulmonary high grade neuroendocrine carcinoma. Four patients (4.7%) had EGFR-mutant lesions, 9 patients (10.6%) had KRAS-mutant lesions, One patient (1%) had a ROS-1 mutant lesion, and 12 patients (14.1%) demonstrated PD-L1 expression in > 50% of cells via tumor proportion score (TPS). No patient had an ALK mutation. We defined tumor core volume as the enhancing tumor volume + central non-enhancing tumor volume and the whole tumor volume as tumor core volume + tumor edema volume. Mean (SD) whole tumor volume was 105.5 cc ± 63.4. Mean tumor core volume was 22.1 cc ± 18.5. Mean enhancement volume was 15 cc ± 12.2. Mean tumor edema volume was 83.5 cc ± 57.2. Mean whole brain volume was 1393.5 cc ± 159.5. Mean edema-to-tumor core volume ratio was 10.5 ± 33.1. Table 3 presents the mean and median values for the AI-generated volume measurements among the included patients. Table 3 Tumor volume characteristics and their relationships with tumoral edema. The mean and median volumes of the tumor enhancement, tumor itself, the brain, peritumoral edema, and edema to tumor ratio were calculated (with standard deviations and interquartile ranges, respectively). Spearman correlation coefficients for the association of tumor edema with other tumor volume characteristics were calculated. P-values in bold are statistically significant. Volume Mean ± SD Median [IQR] Spearman Correlation Coefficient (p-value) Enhancement, cc 15 ± 12.2 11.5 [6.8, 19.8] 0.25 ( 0.022) Tumor core, cc 22.1 ± 18.5 16 [8.4, 30.7] 0.13 (0.23) Whole tumor, cc 105.5 ± 63.4 100.8 [60.6, 145.4] 0.95 ( < 0.005) Brain, cc 1393.5 ± 159.5 1374.2 [1295.4, 1496.3] 0.07 (0.55) Edema, cc 83.5 ± 57.2 73.9 [36.1, 111.1] Edema to tumor ratio 10.5 ± 33.1 4.4 [1.6, 8.9] Univariate Analysis The correlation of tumor edema volume with enhancing tumor volume was statistically significant with a coefficient of + 0.25 and P = 0.022. Similarly, whole tumor volume displayed a statistically significant correlation with tumor edema, with a coefficient of + 0.95 and P < 0.005. Neither tumor core nor whole brain volume were significantly associated with tumor edema. Table 3 displays Spearman correlation coefficients for the association of tumor edema to other tumor volume characteristics. Patients with cerebellar lesions were associated with significantly lower tumor edema volume (negative regression coefficient, P < 0.005). Similarly, patients with midline lesions were associated with significantly lower tumor edema volume (negative regression coefficient, P < 0.005). Patients with supratentorial tumors were significantly associated with higher tumor edema volumes (positive regression coefficient, P = 0.003). Table 1 demonstrates univariate linear regression analyses to detect associations between each tumor location variable and tumor edema. Age, sex, comorbidities, and molecular markers were not associated with significant differences in tumor edema-to-core ratio. Patients with nausea were associated with significantly lower edema-to-core ratios (negative coefficient, P = 0.012), while patients with confusion or focal deficits were associated with significantly higher edema-to-core ratios (positive coefficients, P = 0.024, and P = 0.001, respectively). None of the other investigated comorbidities demonstrated statistical significance in this comparison. Table 4 demonstrates univariate linear regression analysis to detect associations between each patient-related variable and the volume ratio of tumor edema to core. Table 4 Patient characteristics and the tumor edema to core ratio. Univariate linear regression analysis was completed to detect associations between each patient variable and the ratio of the tumor edema to core. Age of patient, sex of patient, comorbidities, cancer molecular markers, and pre-diagnosis symptoms were evaluated for potential associations with tumor edema. P-values in bold are statistically significant. Variable Regression coefficient 95% Confidence Interval P-value Age -1.06 -2.3 to 0.2 0.1 Female -10.04 -34.9 to 14.8 0.42 Comorbidities Hypertension -15.46 -40.2 to 9.3 0.22 Diabetes -12.49 -50.9 to 25.9 0.52 Cardiac -9.49 -45.1 to 26.1 0.6 Asthma/pulmonary -3.11 -28.1 to 21.8 0.81 Smoking 61.58 -4.3 to 127.5 0.067 Molecular markers EGFR 22.67 -35.7 to 81.1 0.44 ROS-1 -1.18 -41.5 to 39.2 0.95 PD-L1 22.23 -13.1 to 57.5 0.21 Symptoms Headaches 2.89 -22.4 to 28.2 0.82 Nausea -34.69 -61.4 to -7.9 0.012 Memory 31.82 -0.2 to 63.9 0.052 Confusion 30.78 4.2 to 57.4 0.024 Focal deficit 41.07 17.5 to 64.6 0.001 Balance issues -1.53 -26.8 to 23.7 0.9 Seizures 11.59 -21.1 to 44.3 0.48 Cranial nerve palsy 5.24 -77.1 to 87.6 0.9 Multivariable Analysis Table 5 demonstrates multivariable linear regression between tumor edema and enhancing component and tumor core, controlling for the tumor being in the supratentorial compartment. The choice of the supratentorial compartment as the variable to control for is based partly on pre-defined clinical knowledge, as well as it being statistically significant during univariate analysis. Multivariable regression showed tumor enhancement to have a statistically significant regression coefficient of + 1.53 to tumor edema volume, with a P value of 0.003. This was with statistical significance on the controlled variable of supratentorial component, with P = 0.008. However, analysis of tumor core to tumor edema volume did not reach statistical significance, demonstrating that extent of enhancement was a stronger determinant of edema volume rather than the entire tumor core volume. Table 5 Multivariable linear regression between tumor edema and enhancing component and tumor edema and tumor core, controlling for the tumor being in the supratentorial compartment. P-values in bold are statistically significant. Variable Regression coefficient 95% Confidence Interval P-value Core 0.65 -0.01 to 1.3 0.053 Supratentorial 42.3 12.2 to 72.5 0.007 Enhancement 1.53 0.5 to 2.5 0.003 Supratentorial 39.9 10.7 to 69.1 0.008 Discussion Peritumoral edema is a significant contributor of morbidity and mortality in patients with BM. Management of peritumoral edema remains a challenge and assessment of its clinicopathological properties may both further our understanding of its pathogenesis and improve management. In this proof-of-concept study, we demonstrate the use of DLBAI to assist in topographical and clinicopathological pattern detection of BM-associated peritumoral edema. In our pilot study of 84 patients with NSCLC and BM, we found that our deep learning algorithm was able to successfully segment radiographic features into 5 distinct partitions, which we extrapolated to our ETR value. We then incorporated this data into our regression analysis and found that tumors in all locations demonstrated a mean 10:1 ETR and that a cerebellar or midline lesion was associated with a significantly lower ETR. We did not detect any significant correlation between ETR and our clinical or histologic variables including presenting symptoms, comorbidities, and selected molecular markers. One potential explanation for the lower cerebellar ETR found in our analysis may reflect both the cerebellum’s overall volume and its relative paucity of white matter tracts compared to supratentorial regions. White matter tracts are classically characterized by both volumetric and connective features. Connectivity-based definitions have become more prominent given the implementation of diffusion tractography[ 11 ]. Vasogenic edema, which is the primary edema associated with BMs, predominantly affects white matter tracts. In a population-based image segmentation analysis study, Hoogendam et al. demonstrated that white matter comprised 20% of total cerebellar volume in comparison to 51% of the total cerebral volume. However, true volumetric segmentation of white matter in the cerebellum proves to be challenging due to its density within the delicately branching cerebellar folia, leading to subvoxel resolution [ 12 ]. Additionally, this theory does not fully explain the statistically significant lower ETR seen with midline lesions in our analysis. With diffusion tractography axonal configuration (directionality and crossover), packing density, and myelination properties for white matter tracts throughout the brain have been elucidated [ 11 ]. Major midline commissural white matter tracts include the corpus callosum, anterior commissure, and hippocampal commissure of the fornix. The corpus callosum is known to be the largest white matter tract within the human brain, with roughly 200 million axons. Clearly, lower ETR in this location cannot be explained by axonal density of white matter tracts alone but may also be a function of spatial distribution and anatomic constraints. The highest axonal density of the corpus callosum lies midline in the rostrum, genu, body, and splenium and segmentally disperses as it intercalates through bilateral cerebral hemispheres [ 13 ]. Anatomically, its supero-inferior margins are defined by the cingulate gyrus superiorly and the ventricles and thalami inferiorly. This produces a “bottleneck” appearance of the fibers as they traverse between cerebral hemispheres and may place an anatomic constraint on the extent of edema dispersion. Another possible variable that may affect peritumoral edema is longitudinal axonal myelin density. Microscopic studies of the corpus callosum demonstrate a widely heterogenous distribution of fibers and myelin density along the antero-posterior axis. Various techniques have been used in vivo and ex vivo to determine myelin density including fractional anisotropy seen in diffusion imaging, neurite orientation dispersion and density imaging (NODDI), and histopathological staining with Luxol Fast Blue (LFB). The distribution of myelin content represents a continuous decrease from posterior to anterior callosal segments, with highest values found in the visual callosal segments [ 13 ]. Unmyelinated fibers have been found to make up 16% of the genu and under 5% of other callosal areas [ 14 ]. The heterogeneity of myelination within the corpus callosum may have an effect on peritumoral edema dispersion pathways. Assessment and surveillance of intracranial lesions is currently completed by a primary qualitative radiographic assessment by a trained neuro-radiologist. While historically highly reliable, this process is subject to inter- and intra-observer bias given its qualitative nature. While primary CNS oncologic disease such as glioma can be multifocal, it has a higher propensity for solitary or clustered lesions, which can be more reliably and critically evaluated. In contrast, BMs typically consist of multiple lesions in discrete and isolated locations. Each lesion needs to be closely monitored for growth or treatment response as even minute alterations may affect management and have prognostic value. DLBAI has been increasingly investigated for its utility in image classification and segmentation. Deep learning-based techniques, like the convolutional neural network (CNN), have already been applied to detect and segment breast cancer in mammograms and lung cancer in computed tomography [ 15 , 16 ]. Recently, many deep learning approaches have been applied to glioma segmentation with the introduction of the publicly available BraTS dataset, which has been utilized by multiple study groups [ 8 , 10 , 15 , 17 ]. The data has further been extrapolated to form predictions on overall survival based on radiomic features interpreted by deep learning software with high accuracy [ 9 ]. DLBAI has been introduced as a method of BM detection in previous studies, primarily with the use of T1Gd [ 18 – 20 ]. This is a challenging task given the similar morphological properties of BM and other structures, such as blood vessels on T1Gd. More recently, multi-sequence MRI has been used for BM detection and segmentation including CUBE, BRAVO, FLAIR, and 3D black-blood sequences with high accuracy as compared to the previously utilized MPRAGE sequence [ 21 , 22 ]. To our knowledge, there are no previous studies that have investigated the use of DLBAI for both tumor and peri-tumoral edema segmentation in tissue-confirmed BMs with further data extrapolation to detect topographical and clinicopathological associations with the extent of peri-tumoral edema. The use of DLBAI for tumor segmentation involving brain edema can be extended in multiple directions. For prognostic and diagnostic purposes this can include comparison to and differentiation from glioblastoma, which can appear radiographically similar to BM and prediction of tumor growth rate and directionality based on edema patterns in both BM and primary infiltrative brain tumors. Additionally, with further immunohistochemical data for, imaging could provide a non-invasive adjunct for prediction of molecular profile and possibly even primary origin of metastases. For treatment purposes, corticosteroids are the mainstay of treatment for cerebral edema. However, dosing has been primarily practitioner-dependent rather than protocol-based. Qualitative volumetric assessment following accurate segmentation of cerebral edema can assist in protocolizing steroid dosing and titration. Importantly, further characterization of BM-associated cerebral edema will help clinicians guide patients toward available treatment modalities, e.g. surgery versus stereotactic radiosurgery/radiotherapy versus by predicting edema treatment response. In our study, we performed a proof-of-concept “snapshot” of NSCLC patients using DLBAI as an inexpensive, efficient, and non-invasive tool to assess the temporal, topographical, immunohistochemical, and clinical characteristics of BM-associated edema. Our methods are not without their limitations. To create a de novo deep learning-based AI algorithm, it typically requires a large sample size for an adequate learning model. For our proof of concept, we limited our sample size to N = 84 out of our full expansive database of patients with BMs. Because of this, we used an existing model developed from the publicly available BraTS dataset for gliomas that has previously been tested and published by members of our study group. While high grade gliomas and BMs may appear radiographically similar there are often features that are more specific to one versus the other, which may not be fully accounted for in our segmentation model. However, a key feature of DLBAI is that it is highly adaptable with incremental human input to account for radiographic minutia, although this cannot be completely ruled out. In this study we demonstrate the ability of DLBAI to effectively be applied to a BM edema model to detect radiographic and topographic patterns, which can be extrapolated to clinicopathologic correlates. We found that overall, tumors in all locations demonstrated a mean 10:1 ETR and both cerebellar and midline tumor locations were associated with a significantly lower ETR. Methods Patient Selection Consecutive patients with tumors metastatic to the brain seen at our institution between March 2010 – September 2019 were reviewed retrospectively. Institutional review board approval was obtained beforehand from the Institutional Review Board for Health Sciences Research (IRB-HSR), with waiver of individual patient consent due to the observational nature of the study. All methods were carried out in accordance with the IRB-HSR. Patients were included if they underwent surgical resection of newly diagnosed BM confirmed on post-operative tissue analysis to be metastatic NSCLC. There were no specific exclusion criteria, however we selected only the first eighty-four eligible patients out of our entire brain lesion cohort of 680 patients for the purposes of this proof-of-concept study. All metastatic lesions were segmented regardless of resection status. Clinical, histological, and radiological baseline and outcome data were retrospectively collected. These included the following variables: age, sex, comorbidities (hypertension, diabetes mellitus, cardiopulmonary disease, hyperlipidemia), family cancer history, personal cancer history, smoking status and pack years, date of primary cancer diagnosis, primary cancer pathology, molecular markers, date of cranial disease diagnosis, location of cranial disease, prior treatment history, presenting symptoms, pre- and post-operative Eastern Cooperative Oncology Group (ECOG) performance status and Karnofsky Performance Status (KPS) scores, and prior or current steroid use at time of diagnosis. Image Analysis Multi-sequence MR images of each patient were included in the image analysis. A model pre-trained on glioma patients from the publicly available BRAin Tumor Segmentation (BraTS) dataset was used for tumor segmentation. This model was developed and tested by Feng, et al in 2020 using a 3D U-Net with adaptations in the training and testing strategies, network structures, and model parameters for brain tumor segmentation. Their results achieved 9th place in the 2018 BraTS challenge and was further extended in order to demonstrate clinical significance to develop linear models based on radiomic features extracted from the segmentation to predict patient overall survival, which achieved 1st place in the 2018 BraTS challenge. Since gliomas and BMs can appear similarly to one another on imaging as they consist of similar sub-regions, i.e., peritumoral edema, necrotic core, enhancing, and non-enhancing tumor core, this model demonstrated high fidelity of segmentation as verified by a senior neuro-radiologist (SP) [ 9 ]. Unlike gliomas, BMs do not contain a “non-enhancing tumor core” given their typical encapsulated and non-infiltrative nature. For the purposes of our study, we modified our segmentation identifiers as defined above. Figure 1 demonstrates the segmentation overlay for a single lesion (Fig. 1 ). Deep learning protocol 1. Pre-processing The input images were multi-sequence MR images, including T1, T1 with contrast (T1Gd), T2, and fluid-attenuated inversion recovery (FLAIR), which were co-registered to the same anatomical template in pre-processing. Then, input images were resampled to have an isotropic 1 mm3 resolution. Each 3D image was then normalized to 0-mean, unit variance by subtracting the mean value and dividing by the standard deviation. The normalized images of all sequences were then concatenated, resulting in a whole input image size of MxNxPx4, where M,N,P correlate to the whole image dimensions, in the following order: T1, T1Gd, T2, FLAIR. For training dataset, corresponding multi-class ground truth label maps were generated. In deployment (prediction), for subjects that do not have all MR sequences (T1, T1Gd, T2, FLAIR), the missing sequences were replaced by an all-zero voxel array. 2. Patch Extraction 3D U-Net architecture can fully exploit volumetric spatial information. However, using the entire images as the input is limited by GPU memory and is suboptimal in terms of training time and accuracy. Therefore, a patch-based segmentation approach was applied, where smaller patches of size 96x96x96x4 were extracted from each subject. During patch selection, an oversampling of 50% was applied, where half of the training patches were forced to foreground voxels. Furthermore, the data augmentation technique was applied to improve the robustness of the model. 3. Training model using 3D UNet architecture The 3D UNet architecture used for model training comprises 5 encoding and 5 decoding blocks. Each encoding block consists of two consecutive 3D convolutional layers, followed by instance normalization and leaky rectified linear activation layers. For the decoding blocks, symmetric blocks were used with skip-connections from corresponding encoding blocks, with 3D convolutional layers replaced by 3D transposed convolutional layers. Features were concatenated to the de-convolution outputs, and the segmentation map of the input patch were expanded to the multi-class ground truth labels. During training, stochastic gradient descent with polynomial learning rate decay and momentum of 0.99 was used. Deep supervision was applied to the network by computing loss at each decoding block except the bottleneck layer and first decoding block. The final loss was calculated as the sum of the weighted loss (dice and cross-entropy) computed at each decoding block. To train a model capable of segmenting with missing sequences, we implemented a novel technique called “sequence dropout”. Our sequence dropout method randomly drops MRI sequences when forming training inputs to prevent complex co-adaptation between sequences in training. During training, we randomly drop n sequences by replacing them with an all-zero voxel array essentially creating pure background to prevent any unwanted interference with the network. The relative probabilities of dropping n random sequences were pn = 0 = 0.4, pn = 1 = 0.3, pn = 2 = 0.2, and pn = 3 = 0.1, respectively. After determining n, the specific sequences that were dropped were chosen randomly with uniform distribution. 4. Prediction In prediction step, a sliding window approach was used to combat the problem of fitting whole images into memory. Individual patches of size 96x96x96x4 were extracted for prediction and are assembled back together to match the original testing images. For each window, the original image and left-right flipped image were both predicted, and the probability was averaged to get segmentation output. 5. Post-processing In post-processing, a 3D-connectivity analysis was performed on the segmentation output to remove small regions and gaussian smoothing was performed. Data Analysis Baseline patient characteristics are presented as follows: Categorical variables are presented as frequencies and percentages. Continuous variables are presented as means and standard deviations (SD) or medians and interquartile ranges [IQR], depending on data normality. Normality was ascertained for each continuous variable using Wilcoxon’s test. The AI-generated data were all presented as both means and medians. Univariate analysis was conducted as follows: Spearman correlation coefficients were calculated for the association of the AI-generated tumor edema volume to the other tumor volume characteristics, including enhancement area, tumor core, whole tumor, and whole brain. Univariate linear regression analysis was utilized to detect associations between each tumor location variable (lobe, side, compartment, presence of multiple lesions) and tumor edema, as determined by the AI model. Univariate linear regression analysis was also used to detect associations between a pre-determined set of clinical patient variables (age, sex, comorbidities, molecular markers, and individual symptoms) and ratio of tumor edema to tumor core, as determined by the AI model. Multivariable linear regression analysis was conducted between AI-generated tumor edema volume and enhancing component volume, while controlling for the tumor being in the supratentorial compartment. This covariate was selected to be controlled for because it reached statistical significance in univariate analysis. A P < 0.05 was considered statistically significant in all analyses. All statistics were conducted using Stata software (StataCorp 2013, Stata Statistical Software: Release 13; StataCorp LP, College Station, TX). Declarations Additional Information Competing interests statement: The authors declare no competing interests. Author Contribution J.Y., J.K., and K.K. wrote the manuscript. K.K., F.F., and G.S. collected the data. X.F. and S.P. produced and provided the software to implement the DLBAI. B.L. provided immunohistochemical analysis on all specimens. J.Y., U.Y., D.S., and M.M. developed the study idea. D.S. provided a patient pool for our database. S.P. and R.S. reviewed all imaging manually. G.M. analyzed the results and wrote the statistical methods. J.Y., J.K., K.K., and G.S. formatted and edited the manuscript for submission. M.M. oversaw the entire project and edited the manuscript. All authors reviewed the manuscript. Data Availability The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request. References Barajas, R. F. & Cha, S. Metastasis in Adult Brain Tumors. Neuroimaging Clin. N. Am. 26, 601–620 (2016). Cardinal, T. et al. Anatomical and topographical variations in the distribution of brain metastases based on primary cancer origin and molecular subtypes: a systematic review. Neuro-Oncol. Adv. 4, vdab170 (2022). Parker, M. et al. Epidemiological trends, prognostic factors, and survival outcomes of synchronous brain metastases from 2015 to 2019: a population-based study. Neuro-Oncol. Adv. 5, vdad015 (2023). Naresh, G. et al. Assessment of Brain Metastasis at Diagnosis in Non-Small-Cell Lung Cancer: A Prospective Observational Study From North India. JCO Glob. Oncol. 7, 593–601 (2021). Toh, C. H., Siow, T. Y. & Castillo, M. Peritumoral Brain Edema in Metastases May Be Related to Glymphatic Dysfunction. Front. Oncol. 11, 725354 (2021). Bilgin, E. Y., Unal, O. & Ciledag, N. Vasogenic Edema Pattern in Brain Metastasis. J. Coll. Physicians Surg. Pak. 32, 1020–1025 (2022). Strugar, J., Rothbart, D., Harrington, W. & Criscuolo, G. R. Vascular permeability factor in brain metastases: correlation with vasogenic brain edema and tumor angiogenesis. J. Neurosurg. 81, 560–566 (1994). Estienne, T. et al. Deep Learning-Based Concurrent Brain Registration and Tumor Segmentation. Front. Comput. Neurosci. 14, 17 (2020). Feng, X., Tustison, N. J., Patel, S. H. & Meyer, C. H. Brain Tumor Segmentation Using an Ensemble of 3D U-Nets and Overall Survival Prediction Using Radiomic Features. Front. Comput. Neurosci. 14, (2020). Gering, D. et al. Measuring Efficiency of Semi-automated Brain Tumor Segmentation by Simulating User Interaction. Front. Comput. Neurosci. 14, 32 (2020). Bullock, D. N. et al. A taxonomy of the brain’s white matter: twenty-one major tracts for the 21st century. Cereb. Cortex 32, 4524–4548 (2022). Hoogendam, Y. Y. et al. Determinants of cerebellar and cerebral volume in the general elderly population. Neurobiol. Aging 33, 2774–2781 (2012). Friedrich, P. et al. The Relationship Between Axon Density, Myelination, and Fractional Anisotropy in the Human Corpus Callosum. Cereb. Cortex N. Y. N 1991 30, 2042–2056 (2020). Sandrone, S. et al. Mapping myelin in white matter with T1-weighted/T2-weighted maps: discrepancy with histology and other myelin MRI measures. Brain Struct. Funct. 228, 525–535 (2023). Zhang, M. et al. Deep-Learning Detection of Cancer Metastases to the Brain on MRI. J. Magn. Reson. Imaging JMRI 52, 1227–1236 (2020). Firmino, M. et al. Computer-aided detection system for lung cancer in computed tomography scans: Review and future prospects. Biomed. Eng. OnLine 13, 41 (2014). Ranjbarzadeh, R. et al. Brain tumor segmentation based on deep learning and an attention mechanism using MRI multi-modalities brain images. Sci. Rep. 11, 10930 (2021). Pennig, L. et al. Automated Detection and Segmentation of Brain Metastases in Malignant Melanoma: Evaluation of a Dedicated Deep Learning Model. AJNR Am. J. Neuroradiol. 42, 655–662 (2021). Farjam, R., Parmar, H. A., Noll, D. C., Tsien, C. I. & Cao, Y. An approach for computer-aided detection of brain metastases in post-Gd T1-W MRI. Magn. Reson. Imaging 30, 824–836 (2012). Zhao, L.-M. et al. Radiomic-Based MRI for Classification of Solitary Brain Metastases Subtypes From Primary Lymphoma of the Central Nervous System. J. Magn. Reson. Imaging JMRI 57, 227–235 (2023). Grøvik, E. et al. Deep learning enables automatic detection and segmentation of brain metastases on multisequence MRI. J. Magn. Reson. Imaging JMRI 51, 175–182 (2020). Oh, J.-H., Lee, K. M., Kim, H.-G., Yoon, J. T. & Kim, E. J. Deep learning-based detection algorithm for brain metastases on black blood imaging. Sci. Rep. 12, 19503 (2022). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3851661","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":267761724,"identity":"c14f8465-dd76-44d1-92b3-59aa8bd7ad76","order_by":0,"name":"Jonathan Yun","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4UlEQVRIiWNgGAWjYBACxoYDDAYMDDYMbEAOmMnAkEBYSwEDQxoJWkDgAwPDYWQ+AS3MjWcMN/xsO5/YJ3384uGCAjsGfvYcAwIOO2Ns2Nt2O7GNL6fg8AyDZAbJnjcEtZgZ8G67bczGw5NwmMeAmcHgBmFbzH/+3XYOpqWewZ4ILQbGvNsOyLHxsB8AajnMYCBBUMuxAmPZf8lALTwMQC3HeSTOPCvAq8VwxuENhm/O2PHI97A//szzp1qOvz15AwEtJ2DO4AEzePAqBwF5/vYHUCb7A1yKRsEoGAWjYIQDADtLSCXb2zYPAAAAAElFTkSuQmCC","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":true,"prefix":"","firstName":"Jonathan","middleName":"","lastName":"Yun","suffix":""},{"id":267761725,"identity":"f5b38f80-b03a-4f13-8e0c-05ec97bdb304","order_by":1,"name":"Kristina Kurker","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Kristina","middleName":"","lastName":"Kurker","suffix":""},{"id":267761726,"identity":"a2aedd9f-067b-499e-b789-300fc53b52f2","order_by":2,"name":"Georgios Maragkos","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Georgios","middleName":"","lastName":"Maragkos","suffix":""},{"id":267761727,"identity":"9f53909a-78bc-4443-9e31-ca757032aee5","order_by":3,"name":"Jeyan Kumar","email":"","orcid":"","institution":"Mayo Clinic","correspondingAuthor":false,"prefix":"","firstName":"Jeyan","middleName":"","lastName":"Kumar","suffix":""},{"id":267761728,"identity":"ca0b49b1-ae2c-4427-9401-d9e26c597b7d","order_by":4,"name":"Grant Sutherland","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Grant","middleName":"","lastName":"Sutherland","suffix":""},{"id":267761729,"identity":"3c40a1ad-e155-4242-b5d3-5671f8616cef","order_by":5,"name":"Faraz Farzad","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Faraz","middleName":"","lastName":"Farzad","suffix":""},{"id":267761730,"identity":"bd7964f6-3450-45d3-bdba-a7457fbd203d","order_by":6,"name":"Rohit Sukumar","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Rohit","middleName":"","lastName":"Sukumar","suffix":""},{"id":267761731,"identity":"f5608e6a-1a3d-4825-8af9-edf2e84c8027","order_by":7,"name":"Xue Feng","email":"","orcid":"","institution":"University of Virginia","correspondingAuthor":false,"prefix":"","firstName":"Xue","middleName":"","lastName":"Feng","suffix":""},{"id":267761732,"identity":"ddd07838-d0a2-4443-90f5-2bd05e2564c4","order_by":8,"name":"Ulas Yener","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Ulas","middleName":"","lastName":"Yener","suffix":""},{"id":267761733,"identity":"e3f55fe5-d6d5-475b-878f-9b81977f816f","order_by":9,"name":"Sohil Patel","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Sohil","middleName":"","lastName":"Patel","suffix":""},{"id":267761734,"identity":"5f4e4d8f-d46d-4ef9-a997-f838c82beefb","order_by":10,"name":"David Schiff","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Schiff","suffix":""},{"id":267761735,"identity":"6bee832b-c98a-4718-b9d6-635f19c8b5cf","order_by":11,"name":"Beatriz Lopes","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Beatriz","middleName":"","lastName":"Lopes","suffix":""},{"id":267761736,"identity":"20958b5e-0690-41b6-b410-d2dd2aec8644","order_by":12,"name":"Melike Mut","email":"","orcid":"","institution":"University of Virginia School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Melike","middleName":"","lastName":"Mut","suffix":""}],"badges":[],"createdAt":"2024-01-10 22:44:39","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3851661/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3851661/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":49895647,"identity":"feace9d2-f9fb-4b20-8be7-efed0fa47510","added_by":"auto","created_at":"2024-01-19 21:44:12","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":199166,"visible":true,"origin":"","legend":"\u003cp\u003eAI segmented tumor imaging. Shows the AI segmented tumor core (yellow) and edema (red) regions overlayed on the post-contrast T1 images from subject RSNSCLC108. This study has pre- and post-contrast T1, T2 and FLAIR sequences, and the segmentation relies on all sequences for different sub-regions.\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-3851661/v1/4061969199e474ce3bb29e2f.jpeg"},{"id":64725634,"identity":"e104502e-1c79-49a0-b7e1-811822b0c158","added_by":"auto","created_at":"2024-09-18 05:34:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":843700,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3851661/v1/2ca38276-6a25-4a5c-9600-4b1801e3a744.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Metastatic non-small cell lung cancer (NSCLC) and brain edema: a topographical and clinicopathological investigation utilizing deep learning-based artificial intelligence (DLBAI)","fulltext":[{"header":"Introduction","content":"\u003cp\u003eBrain metastases (BMs) are the most common adult intracranial malignancy. BMs will occur in approximately 10\u0026ndash;20% of all adult cancer patients and portend a poor prognosis with an estimated median survival of 5 months for synchronous BMs. The epidemiological data is however limited because, unlike most cancers, there is no systematic nationwide reporting of secondary brain malignancies [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The lung is the most common site of origin for BMs, which are found in up to 20% of patients at the time of initial cancer diagnosis [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Non-small-cell lung cancer (NSCLC) is the most common form of lung cancer and up to 40% of patients with NSCLC will develop BMs during their disease course [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIt is well established that BMs are associated with extensive peritumoral edema, which is a significant contributor to their morbidity and mortality. Symptomatology such as neurological deficit and pain secondary to increased intracranial pressure or mass effect can greatly affect both quality of life and overall survival. Pathogenesis of peritumoral edema in BM remains yet to be fully elucidated but is postulated to be secondary to tumor-induced angiogenesis and subsequent alterations in the surrounding microenvironment [\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eUnderstanding both radiographic and clinicopathologic patterns of peritumoral edema of BMs may further elucidate its underlying pathogenesis. To date, primary radiographic assessment of brain tumors is qualitative in nature, using freehand unidimensional and bi-dimensional measurements of specific areas of interest, which can be susceptible to inter-observer and intra-observer variability. Deep learning-based artificial intelligence (DLBAI) tumor segmentation algorithms have recently gained attention to circumvent the time-consuming and potentially biased manual assessment [\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. In this study, we investigate a series of patients with NSCLC with the use of AI-assisted brain tumor segmentation to detect significant topographic patterns of peritumoral edema of BMs, thus demonstrating a proof-of-concept clinical application of these deep learning algorithms.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eBaseline Characteristics\u003c/h2\u003e \u003cp\u003eA total of 84 patients with a surgically resected brain metastases were included for analysis. There were 46 female (54.1%) and 38 male (45.9%) patients. Mean age was 66.2\u0026thinsp;\u0026plusmn;\u0026thinsp;9.6 years. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e demonstrates tumor location characteristics for the patient cohort. Out of the resected lesions, 32 (38.1%) were in the frontal lobe, 8 (9.5%) in the temporal, 9 (10.7%) in the parietal, 8 (9.5%) in the occipital lobe, and 27 (32.1%) were in the cerebellum. Lesions located in the insula were classified as temporal and lesions located in the basal ganglia and thalamus were classified as frontal. There were no brain stem lesions identified in our cohort. Of the 84 patients, 21 (25%) had left-sided lesions, 36 (42.9%) right-sided, and 27 (32.1%) midline lesions. A total of 64 patients (81%) had at least one supratentorial lesion, and 37 (46.8%) had at least one infratentorial lesion. Thirty-six patients (45.6%) had multiple lesions. Eight patients were undergoing systemic therapy at the time of cranial diagnosis. Four patients were on pembrolizumab, 3 patients were on nivolumab, and 1 patient was on atezolizumab. There were no patients on anti-VEGF therapy at the time of diagnosis of their intracranial disease.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatient characteristics. For each patient, age, sex, comorbidities, molecular markers of their cancer, and pre-diagnosis symptoms were recorded and compared.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePatients,\u003c/p\u003e \u003cp\u003en (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e46 (54.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComorbidities\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertension\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38 (44.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 (11.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCardiac\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12 (14.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAsthma/pulmonary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38 (44.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSmoking\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e82 (96.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolecular markers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEGFR mutant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4 (4.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKRAS mutant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9 (10.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePD-L1 expression\u0026thinsp;\u0026gt;\u0026thinsp;50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12 (14.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eROS-1 mutant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1 (1.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eALK mutant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSymptoms\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeadaches\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e37 (44.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNausea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24 (28.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMemory\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15 (17.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConfusion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25 (29.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFocal deficit\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38 (45.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBalance issues\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38 (45.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSeizures\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15 (17.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCranial nerve palsy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (2.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (in years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e66.2\u0026thinsp;\u0026plusmn;\u0026thinsp;9.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eRegarding patient-specific comorbidities, 38 patients (44.7%) had hypertension, 10 patients (11.8%) had diabetes, 12 patients (14.1%) had cardiac comorbidities, 38 patients (44.7%) had asthma or other pulmonary past medical history, and 82 patients (96.5%) were past or present smokers. Thirty-seven patients (44.1%) had headaches, 24 patients (28.6%) complained of nausea, 15 patients (17.9%) had memory issues, 25 patients (29.8%) had confusion, 38 patients (45.2%) had a focal deficit, another 38 patients (45.2%) had balance issues, 15 patients (17.9%) had seizures, and 2 patients (2.4%) had a cranial nerve palsy. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents patient-specific baseline characteristics.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTumor location characteristics and relationships with tumoral edema. The total number of patients and percent of patients was calculated for which lobes the metastatic lesions were located, sidedness of their location, whether they were supratentorial or infratentorial, and whether the patient had multiple metastases. Univariate linear regression analysis was conducted to detect associations between each location variable and tumor edema. The lobe the tumors were in, sidedness of the location, compartment, and whether there were multiple metastases were evaluated for potential associations with tumor edema. P-values in bold are statistically significant.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePatients\u003c/p\u003e \u003cp\u003en (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRegression\u003c/p\u003e \u003cp\u003eCoefficient to Edema Volume\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95% Confidence\u003c/p\u003e \u003cp\u003eInterval\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eP-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLocation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFrontal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32 (38.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e21.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-3.8 to 47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.095\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTemporal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8 (9.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e29.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-12.4 to 72.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.16\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eParietal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9 (10.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-30.2 to 50.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOccipital\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8 (9.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e34.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-7.8 to 76.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.11\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCerebellar\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e27 (32.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-53.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-77.5 to -29.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.005\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSide\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLeft\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21 (25)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e29.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8 to 57.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.044\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e36 (42.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e25.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.4 to 50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.046\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMidline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e27 (32.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-53.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-77.5 to -29.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.005\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCompartment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSupratentorial\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e64 (81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e46.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e16.2 to 77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.003\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInfratentorial\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e37 (46.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-24.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-49.3 to 0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMultiple\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e36 (45.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-7.7 to 42.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.17\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eRegarding tumor histology, 61 patients (73%) had adenocarcinoma, 12 patients (14%) had squamous cell carcinoma, 7 patients (8%) had poorly differentiated NSCLC, 2 patients (2%) had adenosquamous carcinoma, and 2 patients (2%) had pulmonary high grade neuroendocrine carcinoma. Four patients (4.7%) had EGFR-mutant lesions, 9 patients (10.6%) had KRAS-mutant lesions, One patient (1%) had a ROS-1 mutant lesion, and 12 patients (14.1%) demonstrated PD-L1 expression in \u0026gt;\u0026thinsp;50% of cells via tumor proportion score (TPS). No patient had an ALK mutation.\u003c/p\u003e \u003cp\u003eWe defined tumor core volume as the enhancing tumor volume\u0026thinsp;+\u0026thinsp;central non-enhancing tumor volume and the whole tumor volume as tumor core volume\u0026thinsp;+\u0026thinsp;tumor edema volume. Mean (SD) whole tumor volume was 105.5 cc\u0026thinsp;\u0026plusmn;\u0026thinsp;63.4. Mean tumor core volume was 22.1 cc\u0026thinsp;\u0026plusmn;\u0026thinsp;18.5. Mean enhancement volume was 15 cc\u0026thinsp;\u0026plusmn;\u0026thinsp;12.2. Mean tumor edema volume was 83.5 cc\u0026thinsp;\u0026plusmn;\u0026thinsp;57.2. Mean whole brain volume was 1393.5 cc\u0026thinsp;\u0026plusmn;\u0026thinsp;159.5. Mean edema-to-tumor core volume ratio was 10.5\u0026thinsp;\u0026plusmn;\u0026thinsp;33.1. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the mean and median values for the AI-generated volume measurements among the included patients.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTumor volume characteristics and their relationships with tumoral edema. The mean and median volumes of the tumor enhancement, tumor itself, the brain, peritumoral edema, and edema to tumor ratio were calculated (with standard deviations and interquartile ranges, respectively). Spearman correlation coefficients for the association of tumor edema with other tumor volume characteristics were calculated. P-values in bold are statistically significant.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVolume\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMedian [IQR]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpearman Correlation Coefficient (p-value)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhancement, cc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e15\u0026thinsp;\u0026plusmn;\u0026thinsp;12.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11.5 [6.8, 19.8]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.25 (\u003cb\u003e0.022)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTumor core, cc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e22.1\u0026thinsp;\u0026plusmn;\u0026thinsp;18.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16 [8.4, 30.7]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.13 (0.23)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWhole tumor, cc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e105.5\u0026thinsp;\u0026plusmn;\u0026thinsp;63.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e100.8 [60.6, 145.4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.95 (\u003cb\u003e\u0026lt;\u0026thinsp;0.005)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBrain, cc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e1393.5\u0026thinsp;\u0026plusmn;\u0026thinsp;159.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1374.2 [1295.4, 1496.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.07 (0.55)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEdema, cc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e83.5\u0026thinsp;\u0026plusmn;\u0026thinsp;57.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e73.9 [36.1, 111.1]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEdema to tumor ratio\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e10.5\u0026thinsp;\u0026plusmn;\u0026thinsp;33.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4.4 [1.6, 8.9]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eUnivariate Analysis\u003c/h2\u003e \u003cp\u003eThe correlation of tumor edema volume with enhancing tumor volume was statistically significant with a coefficient of +\u0026thinsp;0.25 and P\u0026thinsp;=\u0026thinsp;0.022. Similarly, whole tumor volume displayed a statistically significant correlation with tumor edema, with a coefficient of +\u0026thinsp;0.95 and P\u0026thinsp;\u0026lt;\u0026thinsp;0.005. Neither tumor core nor whole brain volume were significantly associated with tumor edema. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e displays Spearman correlation coefficients for the association of tumor edema to other tumor volume characteristics.\u003c/p\u003e \u003cp\u003ePatients with cerebellar lesions were associated with significantly lower tumor edema volume (negative regression coefficient, P\u0026thinsp;\u0026lt;\u0026thinsp;0.005). Similarly, patients with midline lesions were associated with significantly lower tumor edema volume (negative regression coefficient, P\u0026thinsp;\u0026lt;\u0026thinsp;0.005). Patients with supratentorial tumors were significantly associated with higher tumor edema volumes (positive regression coefficient, P\u0026thinsp;=\u0026thinsp;0.003). Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e demonstrates univariate linear regression analyses to detect associations between each tumor location variable and tumor edema.\u003c/p\u003e \u003cp\u003eAge, sex, comorbidities, and molecular markers were not associated with significant differences in tumor edema-to-core ratio. Patients with nausea were associated with significantly lower edema-to-core ratios (negative coefficient, P\u0026thinsp;=\u0026thinsp;0.012), while patients with confusion or focal deficits were associated with significantly higher edema-to-core ratios (positive coefficients, P\u0026thinsp;=\u0026thinsp;0.024, and P\u0026thinsp;=\u0026thinsp;0.001, respectively). None of the other investigated comorbidities demonstrated statistical significance in this comparison. Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e demonstrates univariate linear regression analysis to detect associations between each patient-related variable and the volume ratio of tumor edema to core.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatient characteristics and the tumor edema to core ratio. Univariate linear regression analysis was completed to detect associations between each patient variable and the ratio of the tumor edema to core. Age of patient, sex of patient, comorbidities, cancer molecular markers, and pre-diagnosis symptoms were evaluated for potential associations with tumor edema. P-values in bold are statistically significant.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRegression coefficient\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e95% Confidence Interval\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-1.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-2.3 to 0.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-10.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-34.9 to 14.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComorbidities\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertension\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-15.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-40.2 to 9.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.22\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-12.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-50.9 to 25.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCardiac\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-9.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-45.1 to 26.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAsthma/pulmonary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-3.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-28.1 to 21.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSmoking\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e61.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-4.3 to 127.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.067\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolecular markers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEGFR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-35.7 to 81.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eROS-1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-1.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-41.5 to 39.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePD-L1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-13.1 to 57.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.21\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSymptoms\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeadaches\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-22.4 to 28.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNausea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-34.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-61.4 to -7.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.012\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMemory\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e31.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-0.2 to 63.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.052\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConfusion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e30.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2 to 57.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.024\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFocal deficit\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e41.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17.5 to 64.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.001\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBalance issues\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-1.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-26.8 to 23.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSeizures\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e11.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-21.1 to 44.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCranial nerve palsy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-77.1 to 87.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eMultivariable Analysis\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e demonstrates multivariable linear regression between tumor edema and enhancing component and tumor core, controlling for the tumor being in the supratentorial compartment. The choice of the supratentorial compartment as the variable to control for is based partly on pre-defined clinical knowledge, as well as it being statistically significant during univariate analysis. Multivariable regression showed tumor enhancement to have a statistically significant regression coefficient of +\u0026thinsp;1.53 to tumor edema volume, with a P value of 0.003. This was with statistical significance on the controlled variable of supratentorial component, with P\u0026thinsp;=\u0026thinsp;0.008. However, analysis of tumor core to tumor edema volume did not reach statistical significance, demonstrating that extent of enhancement was a stronger determinant of edema volume rather than the entire tumor core volume.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMultivariable linear regression between tumor edema and enhancing component and tumor edema and tumor core, controlling for the tumor being in the supratentorial compartment. P-values in bold are statistically significant.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRegression coefficient\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e95% Confidence Interval\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCore\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-0.01 to 1.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.053\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSupratentorial\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e42.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.2 to 72.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.007\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhancement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5 to 2.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.003\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSupratentorial\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e39.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10.7 to 69.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.008\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003ePeritumoral edema is a significant contributor of morbidity and mortality in patients with BM. Management of peritumoral edema remains a challenge and assessment of its clinicopathological properties may both further our understanding of its pathogenesis and improve management. In this proof-of-concept study, we demonstrate the use of DLBAI to assist in topographical and clinicopathological pattern detection of BM-associated peritumoral edema. In our pilot study of 84 patients with NSCLC and BM, we found that our deep learning algorithm was able to successfully segment radiographic features into 5 distinct partitions, which we extrapolated to our ETR value. We then incorporated this data into our regression analysis and found that tumors in all locations demonstrated a mean 10:1 ETR and that a cerebellar or midline lesion was associated with a significantly lower ETR. We did not detect any significant correlation between ETR and our clinical or histologic variables including presenting symptoms, comorbidities, and selected molecular markers.\u003c/p\u003e \u003cp\u003eOne potential explanation for the lower cerebellar ETR found in our analysis may reflect both the cerebellum\u0026rsquo;s overall volume and its relative paucity of white matter tracts compared to supratentorial regions. White matter tracts are classically characterized by both volumetric and connective features. Connectivity-based definitions have become more prominent given the implementation of diffusion tractography[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Vasogenic edema, which is the primary edema associated with BMs, predominantly affects white matter tracts. In a population-based image segmentation analysis study, Hoogendam et al. demonstrated that white matter comprised 20% of total cerebellar volume in comparison to 51% of the total cerebral volume. However, true volumetric segmentation of white matter in the cerebellum proves to be challenging due to its density within the delicately branching cerebellar folia, leading to subvoxel resolution [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Additionally, this theory does not fully explain the statistically significant lower ETR seen with midline lesions in our analysis. With diffusion tractography axonal configuration (directionality and crossover), packing density, and myelination properties for white matter tracts throughout the brain have been elucidated [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Major midline commissural white matter tracts include the corpus callosum, anterior commissure, and hippocampal commissure of the fornix. The corpus callosum is known to be the largest white matter tract within the human brain, with roughly 200\u0026nbsp;million axons. Clearly, lower ETR in this location cannot be explained by axonal density of white matter tracts alone but may also be a function of spatial distribution and anatomic constraints. The highest axonal density of the corpus callosum lies midline in the rostrum, genu, body, and splenium and segmentally disperses as it intercalates through bilateral cerebral hemispheres [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Anatomically, its supero-inferior margins are defined by the cingulate gyrus superiorly and the ventricles and thalami inferiorly. This produces a \u0026ldquo;bottleneck\u0026rdquo; appearance of the fibers as they traverse between cerebral hemispheres and may place an anatomic constraint on the extent of edema dispersion.\u003c/p\u003e \u003cp\u003eAnother possible variable that may affect peritumoral edema is longitudinal axonal myelin density. Microscopic studies of the corpus callosum demonstrate a widely heterogenous distribution of fibers and myelin density along the antero-posterior axis. Various techniques have been used in vivo and ex vivo to determine myelin density including fractional anisotropy seen in diffusion imaging, neurite orientation dispersion and density imaging (NODDI), and histopathological staining with Luxol Fast Blue (LFB). The distribution of myelin content represents a continuous decrease from posterior to anterior callosal segments, with highest values found in the visual callosal segments [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Unmyelinated fibers have been found to make up 16% of the genu and under 5% of other callosal areas [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The heterogeneity of myelination within the corpus callosum may have an effect on peritumoral edema dispersion pathways.\u003c/p\u003e \u003cp\u003eAssessment and surveillance of intracranial lesions is currently completed by a primary qualitative radiographic assessment by a trained neuro-radiologist. While historically highly reliable, this process is subject to inter- and intra-observer bias given its qualitative nature. While primary CNS oncologic disease such as glioma can be multifocal, it has a higher propensity for solitary or clustered lesions, which can be more reliably and critically evaluated. In contrast, BMs typically consist of multiple lesions in discrete and isolated locations. Each lesion needs to be closely monitored for growth or treatment response as even minute alterations may affect management and have prognostic value.\u003c/p\u003e \u003cp\u003eDLBAI has been increasingly investigated for its utility in image classification and segmentation. Deep learning-based techniques, like the convolutional neural network (CNN), have already been applied to detect and segment breast cancer in mammograms and lung cancer in computed tomography [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Recently, many deep learning approaches have been applied to glioma segmentation with the introduction of the publicly available BraTS dataset, which has been utilized by multiple study groups [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The data has further been extrapolated to form predictions on overall survival based on radiomic features interpreted by deep learning software with high accuracy [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDLBAI has been introduced as a method of BM detection in previous studies, primarily with the use of T1Gd [\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. This is a challenging task given the similar morphological properties of BM and other structures, such as blood vessels on T1Gd. More recently, multi-sequence MRI has been used for BM detection and segmentation including CUBE, BRAVO, FLAIR, and 3D black-blood sequences with high accuracy as compared to the previously utilized MPRAGE sequence [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. To our knowledge, there are no previous studies that have investigated the use of DLBAI for both tumor and peri-tumoral edema segmentation in tissue-confirmed BMs with further data extrapolation to detect topographical and clinicopathological associations with the extent of peri-tumoral edema.\u003c/p\u003e \u003cp\u003eThe use of DLBAI for tumor segmentation involving brain edema can be extended in multiple directions. For prognostic and diagnostic purposes this can include comparison to and differentiation from glioblastoma, which can appear radiographically similar to BM and prediction of tumor growth rate and directionality based on edema patterns in both BM and primary infiltrative brain tumors. Additionally, with further immunohistochemical data for, imaging could provide a non-invasive adjunct for prediction of molecular profile and possibly even primary origin of metastases. For treatment purposes, corticosteroids are the mainstay of treatment for cerebral edema. However, dosing has been primarily practitioner-dependent rather than protocol-based. Qualitative volumetric assessment following accurate segmentation of cerebral edema can assist in protocolizing steroid dosing and titration. Importantly, further characterization of BM-associated cerebral edema will help clinicians guide patients toward available treatment modalities, e.g. surgery versus stereotactic radiosurgery/radiotherapy versus by predicting edema treatment response. In our study, we performed a proof-of-concept \u0026ldquo;snapshot\u0026rdquo; of NSCLC patients using DLBAI as an inexpensive, efficient, and non-invasive tool to assess the temporal, topographical, immunohistochemical, and clinical characteristics of BM-associated edema.\u003c/p\u003e \u003cp\u003eOur methods are not without their limitations. To create a de novo deep learning-based AI algorithm, it typically requires a large sample size for an adequate learning model. For our proof of concept, we limited our sample size to N\u0026thinsp;=\u0026thinsp;84 out of our full expansive database of patients with BMs. Because of this, we used an existing model developed from the publicly available BraTS dataset for gliomas that has previously been tested and published by members of our study group. While high grade gliomas and BMs may appear radiographically similar there are often features that are more specific to one versus the other, which may not be fully accounted for in our segmentation model. However, a key feature of DLBAI is that it is highly adaptable with incremental human input to account for radiographic minutia, although this cannot be completely ruled out.\u003c/p\u003e \u003cp\u003eIn this study we demonstrate the ability of DLBAI to effectively be applied to a BM edema model to detect radiographic and topographic patterns, which can be extrapolated to clinicopathologic correlates. We found that overall, tumors in all locations demonstrated a mean 10:1 ETR and both cerebellar and midline tumor locations were associated with a significantly lower ETR.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003ePatient Selection\u003c/h2\u003e \u003cp\u003eConsecutive patients with tumors metastatic to the brain seen at our institution between March 2010 \u0026ndash; September 2019 were reviewed retrospectively. Institutional review board approval was obtained beforehand from the Institutional Review Board for Health Sciences Research (IRB-HSR), with waiver of individual patient consent due to the observational nature of the study. All methods were carried out in accordance with the IRB-HSR. Patients were included if they underwent surgical resection of newly diagnosed BM confirmed on post-operative tissue analysis to be metastatic NSCLC. There were no specific exclusion criteria, however we selected only the first eighty-four eligible patients out of our entire brain lesion cohort of 680 patients for the purposes of this proof-of-concept study. All metastatic lesions were segmented regardless of resection status. Clinical, histological, and radiological baseline and outcome data were retrospectively collected.\u003c/p\u003e \u003cp\u003eThese included the following variables: age, sex, comorbidities (hypertension, diabetes mellitus, cardiopulmonary disease, hyperlipidemia), family cancer history, personal cancer history, smoking status and pack years, date of primary cancer diagnosis, primary cancer pathology, molecular markers, date of cranial disease diagnosis, location of cranial disease, prior treatment history, presenting symptoms, pre- and post-operative Eastern Cooperative Oncology Group (ECOG) performance status and Karnofsky Performance Status (KPS) scores, and prior or current steroid use at time of diagnosis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eImage Analysis\u003c/h2\u003e \u003cp\u003eMulti-sequence MR images of each patient were included in the image analysis. A model pre-trained on glioma patients from the publicly available BRAin Tumor Segmentation (BraTS) dataset was used for tumor segmentation. This model was developed and tested by Feng, et al in 2020 using a 3D U-Net with adaptations in the training and testing strategies, network structures, and model parameters for brain tumor segmentation. Their results achieved 9th place in the 2018 BraTS challenge and was further extended in order to demonstrate clinical significance to develop linear models based on radiomic features extracted from the segmentation to predict patient overall survival, which achieved 1st place in the 2018 BraTS challenge. Since gliomas and BMs can appear similarly to one another on imaging as they consist of similar sub-regions, i.e., peritumoral edema, necrotic core, enhancing, and non-enhancing tumor core, this model demonstrated high fidelity of segmentation as verified by a senior neuro-radiologist (SP) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Unlike gliomas, BMs do not contain a \u0026ldquo;non-enhancing tumor core\u0026rdquo; given their typical encapsulated and non-infiltrative nature. For the purposes of our study, we modified our segmentation identifiers as defined above. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e demonstrates the segmentation overlay for a single lesion (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eDeep learning protocol\u003c/b\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e1. Pre-processing\u003c/h2\u003e \u003cp\u003eThe input images were multi-sequence MR images, including T1, T1 with contrast (T1Gd), T2, and fluid-attenuated inversion recovery (FLAIR), which were co-registered to the same anatomical template in pre-processing. Then, input images were resampled to have an isotropic 1 mm3 resolution. Each 3D image was then normalized to 0-mean, unit variance by subtracting the mean value and dividing by the standard deviation. The normalized images of all sequences were then concatenated, resulting in a whole input image size of MxNxPx4, where M,N,P correlate to the whole image dimensions, in the following order: T1, T1Gd, T2, FLAIR. For training dataset, corresponding multi-class ground truth label maps were generated. In deployment (prediction), for subjects that do not have all MR sequences (T1, T1Gd, T2, FLAIR), the missing sequences were replaced by an all-zero voxel array.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2. Patch Extraction\u003c/h2\u003e \u003cp\u003e3D U-Net architecture can fully exploit volumetric spatial information. However, using the entire images as the input is limited by GPU memory and is suboptimal in terms of training time and accuracy. Therefore, a patch-based segmentation approach was applied, where smaller patches of size 96x96x96x4 were extracted from each subject. During patch selection, an oversampling of 50% was applied, where half of the training patches were forced to foreground voxels. Furthermore, the data augmentation technique was applied to improve the robustness of the model.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3. Training model using 3D UNet architecture\u003c/h2\u003e \u003cp\u003eThe 3D UNet architecture used for model training comprises 5 encoding and 5 decoding blocks. Each encoding block consists of two consecutive 3D convolutional layers, followed by instance normalization and leaky rectified linear activation layers. For the decoding blocks, symmetric blocks were used with skip-connections from corresponding encoding blocks, with 3D convolutional layers replaced by 3D transposed convolutional layers. Features were concatenated to the de-convolution outputs, and the segmentation map of the input patch were expanded to the multi-class ground truth labels. During training, stochastic gradient descent with polynomial learning rate decay and momentum of 0.99 was used. Deep supervision was applied to the network by computing loss at each decoding block except the bottleneck layer and first decoding block. The final loss was calculated as the sum of the weighted loss (dice and cross-entropy) computed at each decoding block.\u003c/p\u003e \u003cp\u003eTo train a model capable of segmenting with missing sequences, we implemented a novel technique called \u0026ldquo;sequence dropout\u0026rdquo;. Our sequence dropout method randomly drops MRI sequences when forming training inputs to prevent complex co-adaptation between sequences in training. During training, we randomly drop n sequences by replacing them with an all-zero voxel array essentially creating pure background to prevent any unwanted interference with the network. The relative probabilities of dropping n random sequences were pn\u0026thinsp;=\u0026thinsp;0\u0026thinsp;=\u0026thinsp;0.4, pn\u0026thinsp;=\u0026thinsp;1\u0026thinsp;=\u0026thinsp;0.3, pn\u0026thinsp;=\u0026thinsp;2\u0026thinsp;=\u0026thinsp;0.2, and pn\u0026thinsp;=\u0026thinsp;3\u0026thinsp;=\u0026thinsp;0.1, respectively. After determining n, the specific sequences that were dropped were chosen randomly with uniform distribution.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4. Prediction\u003c/h2\u003e \u003cp\u003eIn prediction step, a sliding window approach was used to combat the problem of fitting whole images into memory. Individual patches of size 96x96x96x4 were extracted for prediction and are assembled back together to match the original testing images. For each window, the original image and left-right flipped image were both predicted, and the probability was averaged to get segmentation output.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e5. Post-processing\u003c/h2\u003e \u003cp\u003eIn post-processing, a 3D-connectivity analysis was performed on the segmentation output to remove small regions and gaussian smoothing was performed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eData Analysis\u003c/h2\u003e \u003cp\u003eBaseline patient characteristics are presented as follows: Categorical variables are presented as frequencies and percentages. Continuous variables are presented as means and standard deviations (SD) or medians and interquartile ranges [IQR], depending on data normality. Normality was ascertained for each continuous variable using Wilcoxon\u0026rsquo;s test. The AI-generated data were all presented as both means and medians. Univariate analysis was conducted as follows: Spearman correlation coefficients were calculated for the association of the AI-generated tumor edema volume to the other tumor volume characteristics, including enhancement area, tumor core, whole tumor, and whole brain. Univariate linear regression analysis was utilized to detect associations between each tumor location variable (lobe, side, compartment, presence of multiple lesions) and tumor edema, as determined by the AI model. Univariate linear regression analysis was also used to detect associations between a pre-determined set of clinical patient variables (age, sex, comorbidities, molecular markers, and individual symptoms) and ratio of tumor edema to tumor core, as determined by the AI model. Multivariable linear regression analysis was conducted between AI-generated tumor edema volume and enhancing component volume, while controlling for the tumor being in the supratentorial compartment. This covariate was selected to be controlled for because it reached statistical significance in univariate analysis. A P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant in all analyses. All statistics were conducted using Stata software (StataCorp 2013, Stata Statistical Software: Release 13; StataCorp LP, College Station, TX).\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003e \u003cb\u003eAdditional Information\u003c/b\u003e \u003c/h2\u003e \u003cp\u003e Competing interests statement: The authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJ.Y., J.K., and K.K. wrote the manuscript. K.K., F.F., and G.S. collected the data. X.F. and S.P. produced and provided the software to implement the DLBAI. B.L. provided immunohistochemical analysis on all specimens. J.Y., U.Y., D.S., and M.M. developed the study idea. D.S. provided a patient pool for our database. S.P. and R.S. reviewed all imaging manually. G.M. analyzed the results and wrote the statistical methods. J.Y., J.K., K.K., and G.S. formatted and edited the manuscript for submission. M.M. oversaw the entire project and edited the manuscript. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e \u003cp\u003eThe datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBarajas, R. F. \u0026amp; Cha, S. Metastasis in Adult Brain Tumors. Neuroimaging Clin. N. Am. 26, 601\u0026ndash;620 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCardinal, T. \u003cem\u003eet al.\u003c/em\u003e Anatomical and topographical variations in the distribution of brain metastases based on primary cancer origin and molecular subtypes: a systematic review. Neuro-Oncol. Adv. 4, vdab170 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eParker, M. \u003cem\u003eet al.\u003c/em\u003e Epidemiological trends, prognostic factors, and survival outcomes of synchronous brain metastases from 2015 to 2019: a population-based study. Neuro-Oncol. Adv. 5, vdad015 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaresh, G. \u003cem\u003eet al.\u003c/em\u003e Assessment of Brain Metastasis at Diagnosis in Non-Small-Cell Lung Cancer: A Prospective Observational Study From North India. JCO Glob. Oncol. 7, 593\u0026ndash;601 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eToh, C. H., Siow, T. Y. \u0026amp; Castillo, M. Peritumoral Brain Edema in Metastases May Be Related to Glymphatic Dysfunction. Front. Oncol. 11, 725354 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBilgin, E. Y., Unal, O. \u0026amp; Ciledag, N. Vasogenic Edema Pattern in Brain Metastasis. J. Coll. Physicians Surg. Pak. 32, 1020\u0026ndash;1025 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStrugar, J., Rothbart, D., Harrington, W. \u0026amp; Criscuolo, G. R. Vascular permeability factor in brain metastases: correlation with vasogenic brain edema and tumor angiogenesis. J. Neurosurg. 81, 560\u0026ndash;566 (1994).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEstienne, T. \u003cem\u003eet al.\u003c/em\u003e Deep Learning-Based Concurrent Brain Registration and Tumor Segmentation. Front. Comput. Neurosci. 14, 17 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeng, X., Tustison, N. J., Patel, S. H. \u0026amp; Meyer, C. H. Brain Tumor Segmentation Using an Ensemble of 3D U-Nets and Overall Survival Prediction Using Radiomic Features. Front. Comput. Neurosci. 14, (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGering, D. \u003cem\u003eet al.\u003c/em\u003e Measuring Efficiency of Semi-automated Brain Tumor Segmentation by Simulating User Interaction. Front. Comput. Neurosci. 14, 32 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBullock, D. N. \u003cem\u003eet al.\u003c/em\u003e A taxonomy of the brain\u0026rsquo;s white matter: twenty-one major tracts for the 21st century. Cereb. Cortex 32, 4524\u0026ndash;4548 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoogendam, Y. Y. \u003cem\u003eet al.\u003c/em\u003e Determinants of cerebellar and cerebral volume in the general elderly population. Neurobiol. Aging 33, 2774\u0026ndash;2781 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFriedrich, P. \u003cem\u003eet al.\u003c/em\u003e The Relationship Between Axon Density, Myelination, and Fractional Anisotropy in the Human Corpus Callosum. Cereb. Cortex N. Y. N 1991 30, 2042\u0026ndash;2056 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSandrone, S. \u003cem\u003eet al.\u003c/em\u003e Mapping myelin in white matter with T1-weighted/T2-weighted maps: discrepancy with histology and other myelin MRI measures. Brain Struct. Funct. 228, 525\u0026ndash;535 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, M. \u003cem\u003eet al.\u003c/em\u003e Deep-Learning Detection of Cancer Metastases to the Brain on MRI. J. Magn. Reson. Imaging JMRI 52, 1227\u0026ndash;1236 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFirmino, M. \u003cem\u003eet al.\u003c/em\u003e Computer-aided detection system for lung cancer in computed tomography scans: Review and future prospects. Biomed. Eng. OnLine 13, 41 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRanjbarzadeh, R. \u003cem\u003eet al.\u003c/em\u003e Brain tumor segmentation based on deep learning and an attention mechanism using MRI multi-modalities brain images. Sci. Rep. 11, 10930 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePennig, L. \u003cem\u003eet al.\u003c/em\u003e Automated Detection and Segmentation of Brain Metastases in Malignant Melanoma: Evaluation of a Dedicated Deep Learning Model. AJNR Am. J. Neuroradiol. 42, 655\u0026ndash;662 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarjam, R., Parmar, H. A., Noll, D. C., Tsien, C. I. \u0026amp; Cao, Y. An approach for computer-aided detection of brain metastases in post-Gd T1-W MRI. Magn. Reson. Imaging 30, 824\u0026ndash;836 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, L.-M. \u003cem\u003eet al.\u003c/em\u003e Radiomic-Based MRI for Classification of Solitary Brain Metastases Subtypes From Primary Lymphoma of the Central Nervous System. J. Magn. Reson. Imaging JMRI 57, 227\u0026ndash;235 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGr\u0026oslash;vik, E. \u003cem\u003eet al.\u003c/em\u003e Deep learning enables automatic detection and segmentation of brain metastases on multisequence MRI. J. Magn. Reson. Imaging JMRI 51, 175\u0026ndash;182 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOh, J.-H., Lee, K. M., Kim, H.-G., Yoon, J. T. \u0026amp; Kim, E. J. Deep learning-based detection algorithm for brain metastases on black blood imaging. Sci. Rep. 12, 19503 (2022).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-3851661/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3851661/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTumor-associated vasogenic brain edema is a well-known contributor of morbidity and mortality in patients with metastatic disease to the brain. It is widely accepted that brain metastases (BM) is associated with extensive edema and can cause increased symptomatology such as pain, neurologic deficit, and elevated intracranial pressure depending on extent and location. We present a proof-of-concept retrospective analysis utilizing DLBAI to segment and detect radiological and topographical patterns of peritumoral edema and assess for clinicopathological correlates in 84 patients with NSCLC and BM who underwent surgical resection and were not previously on steroids. We found that overall, tumors in all locations demonstrated a mean 10:1 edema to tumor ratio (ETR) and an occipital tumor location was associated with a significantly elevated ETR. Within our cohort there were no other factors that were significantly associated with ETR. This study demonstrates a proof-of-concept that DLBAI is an efficient and accurate method of radiographic analysis that can be applied to detect and potentially predict clinicopathological data and prognostic determinants. Clinically, we demonstrate that NSCLC is associated with significant peritumoral edema and that topographical factors may be associated with increased extent of edema.\u003c/p\u003e","manuscriptTitle":"Metastatic non-small cell lung cancer (NSCLC) and brain edema: a topographical and clinicopathological investigation utilizing deep learning-based artificial intelligence (DLBAI)","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-19 21:44:07","doi":"10.21203/rs.3.rs-3851661/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ac72c8ce-e2c5-4a65-849e-09ce5f29f6d9","owner":[],"postedDate":"January 19th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":28212655,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":28212656,"name":"Biological sciences/Neuroscience"},{"id":28212657,"name":"Health sciences/Medical research"},{"id":28212658,"name":"Health sciences/Neurology"},{"id":28212659,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2024-09-18T05:18:07+00:00","versionOfRecord":[],"versionCreatedAt":"2024-01-19 21:44:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3851661","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3851661","identity":"rs-3851661","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}