Search

BackgroundOne in eight men will be affected by prostate cancer (PCa) in their lives. While the current clinical standard prognostic marker for PCa is the Gleason score, it is subject to interreviewer variability. This study compares two machine learning methods for discriminating between high- and low-grade PCa on histology from 47 PCa patients.MethodsDigitized slides were annotated by a GU fellowship-trained pathologist. High-resolution tiles were extracted from annotated and unlabeled tissue. Glands were segmented and pathomic features were calculated and averaged across each patient. Patients were separated into a training set of 31 patients (Cohort A, n=9345 tiles) and a testing cohort of 16 patients (Cohort B, n=4375 tiles). Tiles from Cohort A were used to train a compact classification ensemble model and a ResNet model to discriminate tumor and were compared to pathologist annotations.ResultsThe ensemble and ResNet models had overall accuracies of 89% and 88%, respectively. The ResNet model was additionally able to differentiate Gleason patterns on data from Cohort B while the ensemble model was not.ConclusionsOur results suggest that quantitative pathomic features calculated from PCa histology can distinguish regions of cancer; how-ever, texture features captured by deep learning frameworks better differentiate unique Gleason patterns.

BackgroundThis study identified a clinically significant subset of glioma patients with tumor outside of contrast-enhancement present at autopsy, and subsequently developed a method for detecting non-enhancing tumor using radio-pathomic mapping. We tested the hypothesis that autopsy-based radio-pathomic tumor probability maps would be able to non-invasively identify areas of infiltrative tumor beyond traditional imaging signatures.MethodsA total of 159 tissue samples from 65 subjects were aligned to MRI acquired nearest to death for this study. Demographic and survival characteristics for patients with and without tumor beyond the contrast-enhancing margin were computed. An ensemble algorithm was used to predict pixelwise tumor presence from pathological annotations using segmented cellularity (Cell), extracellular fluid (ECF), and cytoplasm (Cyt) density as input (6 train/3 test subjects). A second level of ensemble algorithms were used to predict voxel-wise Cell, ECF, and Cyt on the full dataset (43 train/22 test subjects) using 5-by-5 voxel tiles from T1, T1+C, FLAIR, and ADC as input. The models were then combined to generate non-invasive whole brain maps of tumor probability.ResultsTumor outside of contrast was identified in 41.5 percent of patients, who showed worse survival outcomes (HR=3.90, p<0.001). Tumor probability maps reliably tracked non-enhancing tumor in the test set, external data collected pre-surgery, and longitudinal data to identify treatment-related changes and anticipate recurrence.ConclusionsThis study developed a multi-1 stage model for mapping gliomas using autopsy tissue samples as ground truth, which was able to identify regions of tumor beyond traditional imaging signatures.

Tumor heterogeneity in glioblastoma complicates delineation of active tumor using standard MR imaging. Pseudo-progression following treatment with chemotherapy and radiation (chemoRT) further complicates how tumors appear. T1-weighted subtraction maps (T1S) have been used to better identify subtly enhancing regions containing infiltrative tumor. This study examines the differences in tumor appearance between patients treated with chemoRT compared to a cohort opting out at autopsy, to understand how chemoRT changes contrast enhancement on MRI. Ten patients diagnosed with glioblastoma were recruited for whole brain donation for this study. Three patients received no treatment (chemoRT-), and seven received a combination of chemoRT and additional treatments (chemoRT+), including but not limited to bevacizumab (Bev) and tumor treating fields (TTF). Large tissue samples were taken at autopsy from whole brain samples sliced axially to align with the last clinical MRI using patient-specific 3D-printed slicing jigs. All tissue samples were hematoxylin and eosin (HE) stained and digitized at 40X resolution (27 total samples). The whole slide images (WSI) were annotated to outline regions containing necrosis without pseudopalisading cells, tumor with pseudopalisading necrosis, and infiltrative tumor. T1S were created for each patient by subtracting intensity normalized T1-weighted images from T1 post-contrast images (T1C). The annotated WSIs were aligned and resampled into MRI space using a custom software. A mixed effect model was used to compare the T1S intensity between chemoRT +/- cohorts within each pathological annotation, incorporating a random effect for subject. Mean T1S intensity was greater in untreated subjects when compared to chemoRT-subjects within each pathological annotation, including necrosis without pseudopalisading cells, tumor with pseudopalisading necrosis, and infiltrative tumor (p< 0.001). We show that chemoRT reduces the contrast enhancement in all aspects of pathologically validated tumor compartments, including infiltrative tumor. Future research is needed to examine if other patterns are evident in additional MR sequences.

PURPOSE Glioblastoma is one of the most common and deadly adult brain tumors. Current standard treatment is surgical resection followed by radiation and concomitant chemotherapy (chemoRT). Glioblastoma progression is monitored using MRI, primarily relying on post-contrast T1-weighted imaging (T1C). Unfortunately, tumor invasion is known to extend beyond traditional contrast enhancement. T1-subtraction (T1S) maps have been introduced as a better tumor volume estimate. In this study we compare T1S map derived tumor annotations to a radiologist for identifying histologically confirmed tumor in patients with differing treatment histories at autopsy. METHODS Ten patients with autopsy confirmed glioblastoma and MRI within 1 month of death were recruited for this study. Seven patients received chemoRT combined (chemoRT+) and three patients received no treatment beyond surgery (chemoRT-). Patient’s brains were sliced axially in the same orientation as their final MRI using a patient-specific slicing jig. Large tissue samples were taken, processed, embedded in paraffin, stained for hematoxylin and eosin, and digitized at 40x resolution. Digital images were annotated for infiltrative tumor, pseudopalisading necrosis, and necrosis without palisading cells. T1S and radiologist annotations were created for each patient using their final MRI (mean 18 days prior to death). The annotated histology images were aligned and resampled into MRI space using custom software and the overlap of pathologically confirmed tumor and MRI derived annotations was compared. RESULTS T1S maps alone were significantly better at identifying areas of histologically confirmed tumor in chemoRT+ patients compared to chemoRT- patients (p=0.043). T1S derived annotations overlapped with 52% of histologically confirmed tumor in the chemoRT- patients and 78% in the chemoRT+. The radiologist drawn tumor masks were more accurate in chemoRT+ patients, identifying 61% confirmed tumor (trending, p=0.097, chemoRT+=61%). CONCLUSION These results demonstrate the difficulty of identifying pathologically confirmed tumor outside contrast enhancement in glioblastoma patients, even in the untreated state.

PURPOSE Grade III and IV astrocytomas and glioblastomas are recognized as high-grade gliomas with aggressive heterogenous natures that carry with them a low overall survival (OS). Standard treatment currently involves surgical resection followed by radiation (RT) and concomitant chemotherapy. Despite this standard pathway, there is often disagreement in the optimal timing of radiation initiation post-surgery. In this study, we examined a subset of patients diagnosed with primary high-grade gliomas that underwent radiation following surgery at varying time frames. METHODS 53 patients from our brain cancer bank with surgical diagnosis of high-grade gliomas were retrospectively analyses for this study. All patients were grouped into three cohorts based on the timing of the initiation of radiation post-surgery. Group 1 RT < 4 weeks, Group 2 RT 4-5 weeks, and Group 3 RT > 5 weeks post-surgery. Overall survival was calculated in days from initial surgical resection to time of death. RESULTS Across the three patient cohorts, Group 1 included 9 patients (mean OS 430 days), Group 2 had 18 patients (mean OS 597 days), and Group 3 had 26 patients (mean OS 895 days). Overall survival trended towards a difference between the three radiation timing groups, with Group 3 showing better survival outcomes compared to Group 1 (p=0.07). CONCLUSIONS This current study demonstrates a positive outcome with increased time between surgical intervention and initiation of radiation and concomitant chemotherapy. While we did not correct for additional treatment outside of standard clinical care, we would expect to observe similar trends. Despite these promising results for treatment planning, future research in larger cohorts is necessary to investigate possible molecular markers that may also impact treatment response and overall survival.

PURPOSE Bevacizumab (Bev) treatment for recurrent glioblastoma ceases tumor angiogenesis, which changes tumor appearance on Gd enhanced T1-weighted imaging. However, little is known about how non-angiogenic portions of tumor are affected by Bev. Therefore, this study tested the hypothesis that patients with specific tumor appearance phenotypes outside contrast enhancement, as identified using autopsy tissue-based radio-pathomic maps, would show differences in survival benefit from Bev treatment. METHODS Previously published methods were used to create machine learning models that generate whole-brain maps of cellularity (Cell), extracellular fluid (ECF) and cytoplasm (Cyt) using glioma autopsy tissue aligned to the patient’s last MRI as ground truth (Bobholz et al. 2022). These maps were then generated for an independent dataset of glioblastoma patients (n=80) using baseline imaging acquired prior to treatment. Patients were then graded for qualitative characteristics of the non-enhancing margin, with phenotypes including Well-Circumscribed (WC) patients with no abnormal tissue surrounding the core, Hypercellular Front (HF) patients with areas of hypercellularity extending from the core, Necrotic Front (NF) patients with areas of necrosis extending from the core, and Hybrid Front (HYF) patients with both hypercellular and necrotic presence beyond the enhancing region. Kaplan Meier analysis was then used to compare survival outcomes between patients who did and did not receive Bev treatment within each phenotype to assess differences in treatment efficacy. RESULTS Only patients with NF or HYF showed significant survival increase with bevacizumab treatment (HR=2.35, p=0.02; and HR=2.45, p=0.03, respectively), with no significant or trending difference in survival observed for WC and HF patients. CONCLUSION Radio-pathomic phenotypes identify patients who show the greatest survival benefit from Bev treatment, which could be used to direct clinical decision-making.

PURPOSE Glioblastoma time to recurrence following initial surgery is difficult to predict as it differs widely between patients. This is important as those who recur early have a poorer prognosis and shorter survival time. This study aimed to test the hypothesis that cell density, as defined by a predictive radio-pathomic mapping model, would indicate a more aggressive tumor, and thus carry a lower time to progression. METHODS 18 confirmed glioblastoma patients were included in this study. All patients underwent surgery followed by chemo-radiation, consistent with standard of care. Inclusion criteria also included radiographic recurrence, and autopsy confirmation of recurrent glioblastoma. Three magnetic resonance imaging (MRI) timepoints were investigated: pre- and post-surgery, and tumor recurrence defined by a radiologist. Patients were classified into two groups, early recurrence, as defined by tumor progression in the first 6 months post-surgery MRI (n=9, average 116 days to recurrence), and late recurrence, which included everyone else (n=9, average 283 days to recurrence). Contrast enhancement and FLAIR hyperintensity regions of interest were annotated from the patients’ T1+C and FLAIR scans. Radio-pathomic maps of predicted tumor cellularity were generated from a previously published machine learning model trained to identify tumor pathology using aligned autopsy tissue samples as ground truth to clinical MRI scans. The T1, T1+C, FLAIR, and apparent diffusion coefficient (ADC) images were used as input. Tumor cellularity values were then averaged across the T1+C and FLAIR ROIs. RESULTS Both pre- and post-surgical cell density within contrast enhancement was significantly greater in patients with early recurrence compared to those who recurred later (p ≤0.05). CONCLUSIONS Our results suggest that radio-pathomic maps of cell density can identify early-recurrence in patients prior to treatment. This may help with treatment planning for radiologists, surgeons, and neuro-oncologists which may include more aggressive surgery and more frequent monitoring.

PURPOSE Although tumor is known to exist beyond the contrast-enhancing margin, a lack of access to pathological information beyond enhancement has hindered the study of non-angiogenic tumor characteristics and their effects on patient prognosis. This study tested the hypothesis that radio-pathomic maps of tumor characteristics developed using autopsy tissue samples can identify phenotypes relevant to patient survival and treatment outcomes. METHODS A dataset of autopsy tissue samples aligned to the MRI as ground truth from 65 glioma patients (training set n=43, test set n=22) was used to train machine learning models that predict and map tumor cellularity (Cell), extracellular fluid (ECF), and cytoplasm (Cyt) density following our previously published methodology (Bobholz et al. 2022). Cell, ECF, and Cyt maps were then generated from the pre-surgical MRI scans from an independent dataset of 80 glioblastoma patients. Each patient was separated into groups based on the appearance of tumor characteristics beyond the contrast-enhancing margin. Group phenotypes included Well-Circumscribed (WC) tumors with no tumor activity beyond contrast enhancement, Necrotic Front (NF) tumors with areas of necrosis extending beyond the tumor margin, Hypercellular Front (HF) tumors with areas of increased cellularity surrounding the tumor, and Hybrid Front (HYF) tumors with characteristics of both HF and NF tumors. A Cox regression was used to assess survival differences between phenotypes, controlling for treatment history. RESULTS Across the 80 patients, 22 were classified as WC, 14 were classified as HF, 24 were classified as NF, and 20 were classified as HYF. NF, HF, and HYF tumors each showed significant/trending reductions in survival when compared to WC tumors (HR=2.02, p=0.03; HR=2.0, p=0.06; and HR=1.75, p=0.09, respectively). CONCLUSION Radio-pathomic phenotypes identify characteristics beyond the contrast-enhancing margin that affect overall survival outcome in glioblastoma.

DICER1-mutant intracranial sarcomas were recently defined by WHO in the 2021 CNS tumor classifications. Currently there is a lack of literature describing the clinical nature of this disease. Here we present a case of primary intracranial sarcoma with DICER1 mutation and unknown etiology. A 44-year-old female presented with aphasia and right-sided weakness. Clinical imaging confirmed the presence of a hemorrhagic mass in the left frontal lobe and superior falx. The patient underwent two surgical resections, but a clear diagnosis was difficult to make due to heterogenous histological features and inconclusive genomic testing. The surgical samples were sent to the National Institute of Health where genetic testing indicated a DICER1 E1813D mutation with high allele frequency (97%), a TP53 mutation with high allele frequency (95%), and a high variant allele frequency NF1 splice site mutation (96%). This resulted in a diagnosis of DICER1-mutant primary intracranial sarcoma. The patient was treated with temozolomide, bevacizumab, and CPT-11. 382 days after the mass was originally identified, the patient died. A detailed analysis of the radiographic images was performed by creating T1-weighted post-contrast enhancement maps (T1C) and T2-weighted fluid attenuated inversion recovery (FLAIR) hyper-intensity maps for each timepoint. We plotted these volumes over time alongside the patient’s treatment history. Large spikes in T1C and FLAIR volume were seen 4 days prior to initiation of temozolomide and 31 days after CPT-11 treatment was ended. At the final clinical MRI, both volumes were about 4 times greater than that of the initial scan. Despite this disease’s aggressive nature, it can be difficult to accurately diagnose, due in part to very few studies investigating the clinical features of this rare tumor. With the addition of this diagnosis to WHO CNS5, future studies should examine the behavior of these tumors, aiding in earlier diagnosis and improvement of treatment guidance.

Current MRI signatures of brain cancer often fail to identify regions of hypercellularity beyond the contrast enhancing region. Therefore, this study used autopsy tissue samples aligned to clinical MRIs in order to quantify the relationship between intensity values and cellularity, as well as to develop a radio-pathomic model to predict cellularity using MRI data. This study used 93 samples collected at autopsy from 44 brain cancer patients. Tissue samples were processed, stained for hematoxylin and eosin (HE) and digitized for nuclei segmentation and cell density calculation. Pre- and post-gadolinium contrast T1-weighted images (T1, T1C), T2 fluid-attenuated inversion recovery (FLAIR) images, and apparent diffusion coefficient (ADC) images calculated from diffusion imaging were collected from each patients’ final acquisition prior to death. In-house software was used to align tissue samples to the FLAIR image via manually defined control points. Mixed effect models were used to assess the relationship between single image intensity and cellularity for each image. An ensemble learner was trained to predict cellularity using 5 by 5 voxel tiles from each image, employing a 2/3-1/3 train-test split for validation. Single image analyses found subtle associations between image intensity and cellularity, with a less pronounced relationship within GBM patients. The radio-pathomic model was able to accurately predict cellularity in the test set (RMSE = 1015 cells/mm2) and identified regions of hypercellularity beyond the contrast enhancing region. We concluded that a radio-pathomic model for cellularity is able to identify regions of hypercellular tumor beyond traditional imaging signatures.

Tumor treating fields (TTFields) are thought to disrupt the cell division process in glioma. This study uses autopsy tissue samples from patients recruited for brain donation to determine whether TTFields treatment duration and usage affects cellularity distributions beyond the gadolinium contrast enhancing region on T1-weighted (T1w) magnetic resonance imaging (MRI). At autopsy, 43 tissue samples from 18 patients who had undergone TTFields treatment (inclusion criteria: >50% compliance and 25-day duration) were collected from brain slices sectioned in-line with slices from the patients’ final MRI prior to death. Nuclei were segmented on the digitized hematoxylin and eosin (HE) stained tissue samples, which were then used to compute cell count across the slides. Tissue was then aligned to the patient’s final MRI scan prior to death using a custom in-house software, where manually defined control points were used to compute a nonlinear transform to warp the tissue to match the MRI. Histogram features including mean, median, 90th percentile, variance, and skewness, were computed from the cellularity distributions within regions outside the traditional tumor margin defined by T1w contrast enhancement for each subject. General linear models were fit to assess the effects of TTFields usage (percent of day) and duration (in days) on each histogram feature, controlling for age, overall survival, and time between TTFields treatment and death, as well as time between MRI acquisition and death. Longer TTFields treatment duration and overall survival was associated with significantly decreased skewness in the non-enhancing cellularity distributions (p< 0.05) while associations with other metrics remained statistically insignificant. These preliminary results in a small patient cohort suggest that TTFields duration may reduce the presence of high cellularity tails in the non-contrast enhancing distributions. Additional research in larger patient populations is warranted to better understand these findings given the number of confounding factors.

Infiltrative glioma beyond contrast enhancement on MRI is often difficult to identify with conventional imaging. In this study, we use large-format autopsy samples aligned to multi-parametric MRI to test the hypothesis that radio-pathomic machine learning models are able to accurately identify areas of infiltrative tumor beyond the contrast enhancing region. At autopsy, 140 tissue samples from 62 brain cancer patients were collected from brain slices sectioned to align with the patients’ last clinical MRI prior to death. Cell, extra-cellular fluid (ECF), and cytoplasm densities were computed from digitized, hematoxylin and eosin-stained samples, and a subset of 20 slides from 9 patients were annotated for tumor presence by a pathologist-trained technician. In-house custom software was used to align the tissue samples to the patients’ last clinical imaging, which included pre- and post-contrast T1, FLAIR, and ADC images. Bagging random forest models were then trained to predict cellularity, ECF, and cytoplasm density using 5-by-5 voxel tiles from each MRI as input. A 2/3-1/3 train-test split was used to validate model generalizability. A naïve Bayes classifier was trained to predict tumor class using cellularity, ECF, and cytoplasm segmentations within the annotation data set, again using a 2/3-1/3 train-test split to validate performance. The random forest models each accurately predicted cellularity, ECF, and cytoplasm density within the test data set, with root-mean-squared error values for each falling within one standard deviation of the ground truth. The histology-based tumor prediction model accurately predicted tumor, with a test set ROC AUC of 0.86. When using whole brain cellularity, ECF, and cytoplasm predictions from the random forest models as inputs for the naïve Bayes classifier, tumor probability maps identified regions of infiltrative tumor beyond contrast enhancement. Our results suggest that radio-pathomic maps of tumor probability accurately identify regions of infiltrative tumor beyond currently accepted MRI signatures.

e14035 Background: With an average overall survival of 12-18 months, glioblastoma has a particularly grim diagnosis. Standard treatment of glioblastoma, following detection on MRI, is surgical resection followed by radiation therapy and chemotherapy and is monitored through MR imaging. Glioblastoma has a unique heterogenous nature that complicates visualization of subtly enhancing tumor. This study used autopsy tissue samples taken from glioblastoma patients with varying treatment, to examine the effects of treatment on cell density within regions of contrast enhancement, using T1-weighted subtraction maps (T1S) from the last MR images to death. Methods: Eight patients diagnosed with glioblastoma at autopsy were recruited for this study. Two patients had no treatment and six received a combination of chemo-radiation and other treatments, including but not limited to bevacizumab (Bev) and TTFields therapy. At autopsy, whole brain samples were sliced axially aligned to the patient’s final MRI to death. Time between last MRI and death ranged from 4-27 days (mean 16 days). Overall survival (OS) ranged from 4-538 days (mean 307 days). Large tissue samples were taken from regions of suspected tumor or treatment effect, for a total of 18 tissue samples. Tissue samples were processed, H&E stained, and digitized at 40X resolution (Huron Slide Scanner). Cell density (cells/mm2) was calculated using digital histology. T1S were created for each patient by subtracting intensity normalized T1 weighted images from T1 post contrast images (T1C). Digital histology was aligned and resampled into MRI space using manual control point registration. Mixed effect models were used to compare differences in cell density across contrast enhancement (T1SE vs. NE) as well as across treatment groups (treatment vs. no treatment). Results: Cellularity was compared across regions of T1S enhancement (T1SE) and non-enhancement (NE) within manually selected regions of interest. Cell density compared between regions of T1SE and NE was not different (p=0.219). Total cell density was increased in patients who had received treatment compared to no treatment in both regions of T1SE and NE (p=0.014). Conclusions: Overall, cell density was increased in patients who had received treatment after diagnosis of glioblastoma. Additional research is needed to examine the extent of treatment’s effect on cellularity of glioblastoma. This work begins to characterize the use of T1S in evaluating glioblastoma tumor burden in patients with varying treatment histories.[Table: see text]