Systems and methods for automated detection of black hole sign for intracerebral hemorrhages

Abstract

Aspects of the disclosure describe a framework incorporating self-supervised learning (SSL) techniques for black hole sign identification on head computed tomography (CT) images. The framework predicts hematoma expansion by retrieving a dataset of CT scans for individuals suffering a stroke, annotating the dataset with scan-level and slice-level annotation, splitting the dataset into a pre-training dataset and a fine-tuning dataset. The framework trains the SSL model in pre-training and fine-tuning phases, finetunes a downstream classifier for slice-level black hole sign identification, and processes, by the SSL model, a diagnostic scan of a patient comprising multiple slices. Probabilities are calculated for each slice associated with black hole sign identification and the framework predicts a probability of hematoma expansion based on an aggregation of the probabilities.

Description

NU2025-046-02 SYSTEMS AND METHODS FOR AUTOMATED DETECTION OF BLACK HOLE SIGN FOR INTRACEREBRAL HEMORRHAGESCROSS REFERENCE TO RELATED APPLICATION^ )

[0001] This application claims priority to provisional Patent Application No. 63 / 790,326 entitled “Systems and Methods for Automated Detection of Black Hole Sign for Intracerebral Hemorrhages” and filed on 4 / 17 / 2025, which is included herein in its entirety.FIELD OF USE

[0002] Aspects of the disclosure relate generally to neuroimaging analysis and more specifically to a framework leveraging self-supervised learning (SSL) techniques for black hole sign (BHS) identification on head computed tomography (CT) images.STATEMENT OF GOVERNMENT INTERESTS

[0003] This invention was made with government support under R01LM013337 and 5R01NS 110779 awarded by the Department of Health and Human Services, National Institutes of Health, and National Institute of Neurological Disorders and Stroke. The government has certain rights in the invention.BACKGROUNDHematoma expansion and prediction

[0004] Intracerebral Hemorrhage (ICH) is the most devastating type of stroke, constituting approximately 10-20% of all strokes and impacting about 80,000 individuals in the US annually. [1,2] Only 15% of individuals diagnosed with ICH patients are functionally independent at follow-up.3Hematoma expansion (HE), growth of the hematoma after the diagnostic scan, predicts disability and death. [4,5] Preventing HE is a biologically plausible objective to improve patient outcomes after ICH. However, the effectiveness of such interventions is hampered by side effects and toxicity resulting from inaccurate patient selection.

[0005] Predicting HE is crucial when identifying patients for targeted treatment. Several risk factors and predictors for HE have been established through the analysis of clinical,6 10imaging, [11-15] and laboratory characteristics. [16-20] Recently, with the advancement in computational phenotyping, machine learning models leveraging clinical, andNU2025-046-02 imaging features for HE have been developed. [21 -23] A significant limitation common to all existing machine learning methods for HE prediction is their dependence on manually curated features and labeling. The reliance on expert physicians to predict HE is a roadblock because expertise has a limited supply, particularly outside of referral centers and during off-hours. Current deep learning methods for identifying HE from head CT scans also suffer from low generalizability due to the missingness of external validation.

[0024] The use of innovative methods, such as self-supervised learning (SSL), minimizes the requirement of manual labeling, achieving better utilization of existing medical imaging datasets. Using SSL could advance the field by minimizing the need for manual labeling.SUMMARY

[0006] The following presents a simplified summary of various aspects described herein. This summary is not an extensive overview and is not intended to identify key or critical elements or to delineate the scope of the claims. The following summary merely presents some concepts in a simplified form as an introductory prelude to the more detailed description provided below. Corresponding apparatus, systems, and computer- readable media are also within the scope of the disclosure.

[0007] To resolve the drawbacks of present methods, the use of innovative methods, such as self-supervised learning (SSL), minimizes the requirement of manual labeling, achieving better utilization of existing medical imaging datasets. Using SSL could advance the field by minimizing the need for manual labeling. Aspects of this disclosure meet an objective to speed up the recognition of which patients are most likely to have HE, prompting rapid hemostatic treatment. Further, aspects relate to development and testing of an algorithm to predict neuroimaging markers known to appearances that predict HE.Methods

[0008] Systems and methods discussed herein provide a novel framework leveraging selfsupervised learning (SSL) techniques for BHS identification on head CT images. In an illustrative example, a ResNet-50 encoder model was pre -trained on over 1.7 million unlabeled head CT images. Layers for binary classification were added on top of the pre-trained model. The resulting model was fine-tuned using the training data and evaluated on the held-out test set to collect AUC and Fl scores. The evaluations wereNU2025-046-02 performed on scan and slice levels. Different panels were run, one using two multicenter datasets for external validation and one including parts of them in the pretraining.Findings

[0009] The illustrative model demonstrated strong performance in identifying BHS when compared with the baseline model. Specifically, the model achieved scan-level AUC scores between 0.75-0.89 and Fl scores between 0.60-0.70. Further, it exhibited robustness and generalizability across an external dataset, achieving a scan-level AUC score of up to 0.85 and an Fl score of up to 0.60, while it performed less well on another dataset with more heterogeneous samples. The negative effects could be mitigated after including parts of the external datasets in the fine-tuning process.Interpretation

[0010] This study introduced a novel framework integrating SSL into medical image classification, particularly on BHS identification from head CT scans. By leveraging SSL, a powerful head CT encoder was trained, eliminating the need for extensive manual annotation, and reducing labor, time, and costs significantly. The resulting pretrained head CT encoder model showed potential to minimize manual annotation, which would significantly reduce labor, time, and costs. After fine-tuning, the framework demonstrated promising performance for a specific downstream task, identifying the BHS to predict HE, upon comprehensive evaluation on diverse datasets. This approach holds promise for enhancing medical image analysis, particularly in scenarios with limited data availability.

[0011] These features, along with many others, are discussed in greater detail below.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present disclosure is described by way of example and not limited in the accompanying figures in which reference numerals indicate similar elements and in which:

[0013] FIG. 1A shows an illustrative image of a black hole sign (BHS) on an axial head CT scan and FIGS. IB and 1C show an illustrative annotation scheme according to one or more aspects of the disclosure;NU2025-046-02

[0014] FIG. 2 shows an illustrative data processing pipeline according to one or more aspects of the disclosure:

[0015] FIG. 3 shows an illustrative framework for prediction of slice-level black hole signs according to one or more aspects of the disclosure;

[0016] FIG. 4 shows an illustrative visualization of positive slices from a set of scans according to one or more aspects of the disclosure;

[0017] FIG. 5 shows a table illustrating case allocation for black hole sign classification according to one or more aspects of the disclosure;

[0018] FIG. 6 shows a table illustrating case statistics of each data source for pre-training according to one or more aspects of the disclosure;

[0019] FIG. 7 shows a table illustrating data allocation in experimental trials according to one or more aspects of the disclosure;

[0020] FIG. 8 shows a table illustrating demographics of patients for black hole sign classification according to one or more aspects of the disclosure;

[0021] FIG. 9 shows tables illustrating slice-level model performance according to one or more aspects of the disclosure;

[0022] FIG. 10 shows tables illustrating scan-level model performance according to one or more aspects of the disclosure;

[0023] FIG. 11 shows tables illustrating performances on the model fine-tuned from ImageNet weights (baseline) according to one or more aspects of the disclosure;

[0024] FIG. 12 shows tables illustrating performances on the model fine-tuned from pretrained weights according to one or more aspects of the disclosure; and

[0025] FIG. 13 shows an illustrative block diagram of a processor platform implementing at least a portion of the self-supervised learning-based black hole sign prediction framework, according to aspects of the present disclosure.DETAILED DESCRIPTION

[0026] In the following description of the various embodiments, reference is made to the accompanying drawings, which form a part hereof, and in which is shown by way of illustration various embodiments in which aspects of the disclosure can be practiced. It is to be understood that other embodiments can be utilized and structural and functionalNU2025-046-02 modifications can be made without departing from the scope of the present disclosure. Aspects of the disclosure are capable of other embodiments and of being practiced or being carried out in various ways. In addition, it is to be understood that the phraseology and terminology used herein are for the purpose of description and should not be regarded as limiting. Rather, the phrases and terms used herein are to be given their broadest interpretation and meaning.Background

[0027] Intracerebral Hemorrhage (ICH) is a devastating form of stroke. Hematoma expansion (HE), growth of the hematoma on interval scans, predicts death and disability. Accurate prediction of HE is crucial for targeted interventions to improve patient outcomes. The black hole sign (BHS) on non-contrast computed tomography (CT) scans is a predictive marker for HE. An automated method to recognize the BHS and predict HE could speed precise patient selection for treatment.Previous Literature

[0028] Preventing hematoma expansion (HE) is an essential objective for patients with intracerebral hemorrhage (ICH). The black hole sign (BHS) visible on non-contrast tomography (CT) was found to be a reliable marker for HE. Previous work focused on HE prediction using machine learning with manually curated features and labeling from CT images. Due to limited expert availability and time constraints, the need for automated methods emerged. Previous deep learning methods predicting HE lacked generalizability because of limited external validation.Key Findings

[0029] The study discussed below proposed a novel framework for medical image classification based on self-supervised learning (SSL). Across different panels it achieved AUC scores between 0.75 and 0.89 and Fl scores between 0.60 and 0.70. Additionally, the systems and methods used with the study outperformed conventional supervised learning frameworks in identifying BHS as a predictor for HE.Knowledge Advancement

[0030] The proposed SSL framework has the potential to minimize manual annotation and labeling efforts. Robust performances across internal and external datasets indicate generalizability for related neuroimaging tasks. While the current implementationNU2025-046-02 limits clinical use, refinement of the framework may further improve diagnostic accuracy of automated predictions.Black Hole Sign

[0031] Humans can recognize several distinct markers on non-contrast neuroimaging that predict future onset of HE. Of several such distinct appearances, the black hole sign (BHS) consistently performs best. [25,26] The BHS is characterized by multiple criteria. First, the BHS comprises a relatively hypo-attenuated area (“black hole”) enclosed within a hyperattenuating hematoma. Second, the black hole may assume various shapes, including round, oval, or rod-like, but should not exhibit any connection to the surrounding brain tissue. Third, the relatively hypo-attenuated area must possess a clearly identifiable border. Fourth, the BHS should comprise a minimum of a twentyeight Hounsfield unit (HU), a measure of radiographic density on CT) contrast between the density of the hematoma and that of the hypo-attenuated area.

[0026] An illustration of the BHS can be found in FIG. 1A. Multiple BHSs may be visible within one scan, and the BHS is smaller than the hematoma.Self-Supervised Machine Learning in Medical Imaging

[0032] Data annotation within medical imaging software incurs significant costs, demanding meticulous data selection and necessitating a substantial investment of time and specialized expertise. SSL represents a machine learning paradigm wherein a model autonomously learns visual features from its own input data, eliminating the reliance on explicitly labeled training data. While it falls under the category of unsupervised learning, SSL places a specific emphasis on utilizing the intrinsic structure or information inherent in the input data to create self-generated supervision signals for training. This approach holds considerable promise for enhancing the robustness of medical imaging models, particularly when labeled data are scarce. By transcending the limitations imposed by a scarcity of labeled data, models trained with SSL can autonomously learn to generate meaningful embeddings or encodings from medical images without the need for manual review or annotation, thus speeding processing time and improving accuracy. SSL facilitates learning from a large volume of images in a specific domain, promoting robust knowledge transfer. Therefore, leveraging SSL techniques in medical imaging analysis presents a promising solution to mitigate the challenges of scarcity in annotated data supply. Nevertheless, we have not identifiedNU2025-046-02 any previous studies specifically employing SSL pre-training on head CT images for cerebrovascular diseases.

[0033] Summarizing from the limited available studies employing SSL for medical imaging, the learning strategies can be characterized as follows: innate relationship learning, generative learning, contrastive learning, and self-prediction. A recent review found contrastive learning to be most effective in radiology and most commonly used in all medical imaging domains.

[0027] The concept behind contrastive SSL is grounded in the assumption that modifications arising from image transformations should preserve the underlying semantic content of the image. Therefore, augmentations generated from the same image constitute positive pairs, whereas those from different images are designated as negative pairs. The model is trained to minimize the distance in the latent space between positive pairs to foster proximity while concurrently maximizing the separation between the negative samples to enhance distinctiveness.

[0034] To address the identified gaps, this novel framework is based on SSL for the effective abstraction of head CT features and subsequently utilized the encoder to execute BHS identification.MethodsData Integration, Annotation, and Processing

[0035] The systems and methods described herein leverage multiple data sources, including in-house data, external datasets, and publicly available data.

[0036] FIG. 1A shows an illustrative image of a black hole sign (BHS) on an axial head CT scan and FIGS. IB and 1C show an illustrative annotation scheme according to one or more aspects of the disclosure. FIG. 1A shows a slice of head CT scan obtained from a patient in the NUB AR dataset. The arrow is pointing at a hypo-attenuated black hole sign within a hyper-attenuated hemorrhagic area. FIGS. IB and 1C demonstrate scanlevel and slice-level annotation for black hole signs.

[0037] BHS annotation was carried out on three datasets, including one in-house dataset (NUBAR) and two external datasets (ITALY, VIZ). The NUBAR dataset contains electronic health records and images for patients with stroke at Northwestern Medicine accumulated over 10 years. The ITALY dataset originated from a Neurology and Stroke Unit in Bologna, Italy (AZ, LB, SG). The VIZ dataset was obtained from Viz.ai an artificial intelligence-powered healthcare software company. It includes patientsNU2025-046-02 with patterns of hemorrhage characteristic of trauma, tumors, and subdural hematoma in addition to hemorrhages consistent with spontaneous intracerebral hemorrhage. Annotation was exclusively performed on these three datasets as they adhere to the inclusion criteria for HE prediction, specifically encompassing diagnostic scans of spontaneous intracerebral hemorrhage (ICH) patients within 12 hours of symptom onset. In addition to scan-level annotation (shown in FIG. IB), slice-level annotation (shown in FIG. 1C) was performed to provide more domain knowledge to the model during supervised training processes. The ground truth labeling was assigned by an expert who did not participate in the construction of the deep learning model. Even though BHS findings can subjectively differ between experts, previous work has demonstrated generally high inter-rater reliability for BHS annotation. [26,28] Axial head CT scans were selected and clipped to the brain window (0-80 HU). A further step was taken to re-slice each set of scans to obtain equidistant slices. The standard slice thickness of scans in the included datasets was 3 mm for infratentorial and 5 mm for supratentorial brain regions. To increase the resolution and because most slices came from the infratentorial region, the samples were re-sliced to a slice thickness of 3 mm.

[0038] FIG. 4 shows an illustrative visualization of positive slices from a set of scans according to one or more aspects of the disclosure. For example, FIG. 4 shows a GradCAM++ visualization of positive slices from a set of illustrative scans from the test set. P comprising probability yielded by the fine-tuned model.

[0039] FIG. 5 shows a table illustrating case allocation for black hole sign classification according to one or more aspects of the disclosure. FIG. 6 shows a table illustrating case statistics of each data source for pre-training according to one or more aspects of the disclosure. FIG. 7 shows a table illustrating data allocation in experimental trials according to one or more aspects of the disclosure. FIG. 8 shows a table illustrating demographics of patients for black hole sign classification according to one or more aspects of the disclosure.

[0040] After annotation, each dataset underwent random training, validation, and testing sets split for both pre-training and fine-tuning. The data split was initiated at the patient level to ensure that slices belonging to a particular patient exclusively appear in one set, avoiding any cross-contamination between training, validation, and testing sets. The respective case counts for each data source are outlined in FIG. 5. The pre-trainingNU2025-046-02 dataset comprises the training sets of the aforementioned datasets, along with non- intracerebral hemorrhage (non-ICH) patients from NUBAR (e.g., cardiac arrest and subarachnoid hemorrhage), an in-house head CT dataset of patients who presented with neurotrauma, and multiple external open-sourced head CT datasets. [29-32] Given that the pre-training process was conducted in a self-supervised learning manner, explicit annotation was unnecessary, and strict data selection criteria were not applicable. Details regarding the number of patients and slices for pre-training are presented in FIG. 6.

[0041] The data processing pipeline and / or procedures are illustrated in FIG. 2. Upon delivery, from a data pipeline 210, data (e.g., the files 220) is received in Digital Imaging and Communications in Medicine (DICOM) format, with each file representing an original slice. To organize scans, essential information, such as timestamps, series details, and study descriptions, is extracted by an extraction engine 230 into a table or data structure. The extracted information may include one or more of DICOM location, Patient ID, Pixel Data, Study Date, Study Time, Acquisition Date, Acquisition Time, Accession Number, Series Description, Window Center, Window Width, Image Type, Body Part Examined, Image Orientation (Patient) Slice Location, Slice Thickness, Instance Number, Series Number Acquisition Number, Rescale Intercept, Rescale Slope, Rescale Type, etc.). The next steps involve selecting one or more axial brain CT scans 240, decompressing each axial brain CT scan by a decompression engine 250, and reslicing them to ensure uniformity by a re-slicing engine 260. The scans are then converted to Neuroimaging Informatics Technology Initiative (NlfTI) format (e.g., the DIfTI file 270), windowed for optimal brain tissue visibility by a windowing engine 280, and finally each slice is resized to 224 by 224 pixels for subsequent analysis by a resizing engine 290.

[0042] Upon receiving data from the Vendor Neutral Archive (VNA) team, the initial format is Digital Imaging and Communications in Medicine (DICOM) format, with each DICOM file representing an original slice. To organize slices from a scan set, pertinent information such as timestamps, series details, and study descriptions was extracted from each DICOM file. Subsequently, we selected axial brain CT scans using the previously extracted information, followed by the decompression of the selected files. Scans were uniformly re-sliced to an equal distance (3 mm) to facilitate a fair comparison across scans, to ensure the inclusion of subtle visual signs, and to align withNU2025-046-02 the assumption in the Neuroimaging Informatics Technology Initiative (NIfTI) format that scans should be equally distant. The subsequent step involves converting each set of scans into a NIfTI file, consolidating all slices from a set into the same file. The images then undergo windowing to enhance the visibility of brain tissue, following which all slices are resized to 224 by 224 pixels analysis.Self-Supervised Learning and Finetuning Framework

[0043] The framework comprising the learning scheme's design is depicted in FIG. 3. An encoder 330 (ResNet50) is pre -trained through the SimCLR framework. Two paired patches 310 are randomly sampled at 320 from each slice image to minimize the contrastive loss between positive and negative pairs. The pre-trained encoder 330 is then fine-tuned on entire slices to make the prediction of slice-level black hole sign (BHS) at 340. The slices with green frames (e.g., the slices 353 where BHS = 0) in the sample scan denote negative for BHS whereas red-framed slices (e.g., the slices 357 where BHS = 1) denote positive for BHS. P1-P18 denotes the probabilities of BHS for the slices 353 and the slices 357 as predicted by the network. The aggregation method performed by an aggregation module 360 utilizes slice-level predictions and generates, as an output 370, a scan-level prediction as a prediction of a black hole sign.

[0044] The model undergoes training in two distinct phases, a self-supervised pre-training phase followed by a supervised learning fine-tuning phase. The primary objective of this training process is to develop an image encoder for head CT and to derive a downstream classifier specifically for the identification of the BHS. In this study, the Simple framework for Contrastive Learning of Visual Representation (SimCLR) was relied on as a common framework to implement self-supervised learning.

[0033] SimCLR is a pioneering contrastive SSL model in the general computer vision domain. It outperformed supervised models on ImageNet using 100 times fewer labels in subsequent fine-tuning, showcasing its capacity for effective feature abstraction. SimCLR is a versatile framework that imposes no constraints on specialized architectures, allowing for the potential use of any visual encoders. For this study, a classic ResNet50 was selected as the encoder based on the performances from small- scale preliminary experiments.

[0045] During the fine-tuning phase, the encoder weights are initialized based on the pretraining results. A downstream classifier, comprised of linear and activation layers, isNU2025-046-02 subsequently fine-tuned to execute slice-level BHS identification. In this stage, the model was fine-tuned upon the manually annotated BHS datasets in a supervised learning manner. The fine-tuned classifier enables the generation of probabilities for any given slice of head CT. Probabilities for all scan slices are aggregated using a maxpooling layer that collects the maximum probability as the final prediction for the patient’s CT scan. The maximum aggregation is used because the presence of BHS on at least one slice determines the overall outcome for the entire scan. The data allocation for each stage, pre-training and fine-tuning, and different trials are shown in FIG. 7. Implementation and evaluation100461 The pre-training was carried out on 8 NVIDIA A100 GPUs. SimCLR learns based on augmentation within batches and, therefore, benefits from large batch sizes, as demonstrated by Chen et al.

[0033] Given the large amount of training data and the computational infrastructure, the largest possible batch size (416) was used. Augmentations applied to generate pre-training pairs include random-size cropping (height / width between 128-200 pixels), random horizontal flipping, and random color jitter (0.8 probability to apply a random augmentation within 0.8, 0.8, 0.8, 0.2 brightness, contrast, saturation, and hue ranges). The learning rate was le’4, and 10 warmup epochs were applied. The model was pre-trained for 75 epochs to ensure convergence based on the monitored loss function. Checkpoints were saved after each epoch. TensorBoard was used to monitor and visualize the pre -training progress. The downstream fine-tuning was implemented on 32G Tesla V100 GPUs. The batch size was 24 per GPU. BCEWithLogitsLoss in PyTorch (Version 2.0.1) was applied for the binary classification task with weights for each class adjusted according to the sample size in training data. The Fl score for the positive class was monitored on the validation set during training. Hyperparameters tuned and their corresponding ranges include learning rate (from le-5to le-4with increment of le’5), weight decay (le1, le-2, or le’3), and batch size (32, 64, or 128). The training was terminated if the monitored Fl score on the validation set failed to improve for 15 consecutive epochs. The area under the receiver operator curve (AUC), precision, recall, and Fl score for the positive class were used for classification evaluation. Fl thresholds were determined by optimizing the hamionic mean of precision and recall under various cutoffs. The thresholds might differ between experiments as independent models were trained for each experiment.NU2025-046-02 GradCAM++

[0034] was used for visual explanations of the model predictions, which also helped to validate initial annotations.

[0047] The Checklist for Artificial Intelligence in Medical Imaging (CLAIM)

[0035] was utilized to ensure comprehensive reporting and is provided in the supplementary material. FindingsDemographics

[0048] Demographics of the cohort for BHS classification are shown in FIG. 8. The demographics of the cohort are reflective of an ICH population.Performance

[0049] The performances of the model are summarized in FIG. 9 and FIG. 10 juxtaposing metrics associated with the baseline encoder and the encoder trained through SSL for comparison. Both external validation (Panel A) and testing performances (Panel B) are presented. While FIG. 9 and FIG. 10 only showcase the major evaluation metrics, detailed slice-level and scan-level performance can be found in FIG. 11 and FIG. 12.

[0050] FIG. 9 showcases the slice-level performances, offering a direct reflection of the model performance before probability aggregation of the calculated probabilities. Panel A displays the performances of models fine-tuned exclusively on NUBAR data, with ITALY and VIZ data utilized for external validation. Panel B presents the performances of models fine-tuned on NUBAR, ITALY, and VIZ data, with all held- out testing sets utilized for testing. Within each panel, the fine-tuning performances are juxtaposed between the baseline encoder and the encoder trained through selfsupervised learning (SSL), with enhancements resulting from the incorporation of the SSL encoder underscored. The SSL encoder improved performance.

[0051] A notable enhancement across all major evaluation metrics is evident when comparing performances fine-tuned from the baseline encoder with those from the SSL encoder. Specifically, the fine-tuned baseline encoder correctly identified zero (ITALY in Panel A) to very few (NUBAR & VIZ) positive cases, as indicated by the recall and positive class Fl score. Those metrics are visibly higher upon utilizing the SSL encoder. Panel A displays the external validation performance on the ITALY and VIZ testing set from the model fine-tuned using NUBAR data. No substantial performance drop was observed from the NUBAR testing set, even for the baseline encoder. ExternalNU2025-046-02 validation performances displayed in Panel A demonstrated comparable robustness as Panel B where all data sources were employed for finetuning. The robustness suggests the validity of the model design and the training strategy. Comparing corresponding cells in Panel A and Panel B, improvements in performance are evident across all testing sets with the incorporation of additional fine-tuning data from each data source.

[0052] The aggregated scan-level performances are presented in FIG. 10. Panel A displays the performances of models fine-tuned exclusively on NUB AR data, with ITALY and VIZ data utilized for external validation. Panel B presents the performances of models finetuned on NUBAR, ITALY, and VIZ data, with all held-out testing sets utilized for testing. Within each panel, the fine-tuning performances are juxtaposed between the baseline encoder and the encoder trained through self-supervised learning (SSL), with enhancements resulting from the incorporation of the SSL encoder underscored. The SSL encoder improved performance. Scan level analysis may be more appropriate for clinical use because the black hole sign is reasonably common (-15%) at the scan level, rather than the slice level (<1%).

[0053] Overall AUC scores of our fine-tuned model ranged from 0.75 to 0.89, with Fl scores from 0.60-0.70 across datasets and panels, except for the external validation with the VIZ dataset showing significantly lower AUC=0.59 and Fl=0.46. In Panel A, the model correctly identified 92.31% of all scans with BHS in the NUBAR cohort (recall) while achieving 44.44% precision. The external validation yielded recall scores of 87.50% and 57.69% for ITALY and VIZ datasets under 46.66% and 38.46% precision. Compared to Panel B, where parts of the external test sets were included in the fine- tuning, the originating model presented more balanced recall and precision, resulting in increased Fl scores. For the NUBAR dataset, 73.08% recall comes to 67.86% precision. Further, the model generates 62.50% for both recall and precision evaluating the ITALY data. VIZ recall and precision scores increased to 77.78% and 53.85%, respectively. Compared to the baseline encoder, we observed consistently higher performances of the SSL encoder across panels and evaluation metrics, except for recall and AUC at external validation with VIZ.

[0054] FIG. 11 shows tables illustrating performances on the model fine-tuned from ImageNet weights (baseline) according to one or more aspects of the disclosure. Panels A, B, C, D show performances of the model fine-tuned only on NUBAR data. ITALY, and VIZ data served as external validations in these cases. Panels E, F, G, H show performancesNU2025-046-02 of the model fine-tuned on NUBAR, ITALY, and VIZ data, all held-out testing sets served as testing. Panels in the left column (A, C, E, G) show slice-level performances, panels in the right column (B, D, F, H) show scan-level performances. Panels A, B, E, F are macro-averaged performances for each condition, panels C, D, G, H are performances on only positive class.

[0055] FIG. 12 shows tables illustrating performances on the model fine-tuned from pretrained weights according to one or more aspects of the disclosure. Panels A, B, C, D show performances of the model fine-tuned only on NUBAR data. ITALY, and VIZ data served as external validations in these cases. Panels E, F, G, H show performances of the model fine-tuned on NUBAR, ITALY, and VIZ data, all held-out testing sets served as testing. Panels in the left column (A, C, E, G) show slice-level performances, panels in the right column (B, D, F, H) show scan-level performances. Panels A, B, E, F are macro-averaged performances for each condition, panels C, D, G, H are performances on only positive class.

[0056] To further enhance the interpretability of the model, a heatmap illustrating the activation of the last feature layer of the framework after fine-tuning for a true positive case is shown in FIG. 4. The heatmap specifically focuses on the slices with positive BHS labels for a patient in the NUBAR cohort. Additionally, corresponding prediction probabilities for each slice are provided. Notably, the model exhibits high confidence in its predictions for such cases as post-fine-tuning, as evidenced by probabilities nearing 1 and precise heatmap localization directly overlapping the BHS.Interpretation

[0057] This study presents a novel framework for medical image classification integrating SSL as a fundamental component. To the best of our knowledge, this work represents the first application of SSL techniques to head CT images for cerebrovascular diseases. The systems and methods described herein demonstrated the usefulness of SSL by solving an important yet difficult clinical problem, BHS identification to predict HE, a predictor of poor patient outcomes. The proposed framework employing an SSL pretraining demonstrated superior performance over end-to-end finetuning of a convolutional neural network (CNN) model. These results suggest SSL is appropriate for identifying visual patterns in moderately sized imaging data sets, an innovative technique for neuroimaging.NU2025-046-02

[0058] Through a comprehensive comparison of model performances, as detailed in FIG. 9 and FIG. 10, the systems and methods described herein showcased the capacity of the trained encoder for head CT image analysis. SSL enabled the encoder to capture critical visual features for BHS prediction during the fine-tuning stage. This finding was reflected in higher performance scores of the fine-tuned classifier using the SSL encoder instead of the baseline encoder. Minor inconsistencies were visible between slice- and scan-level performances and for the external validation with the VIZ dataset. The generally observed drop in AUC from slice -level to scan-level can be attributed to the shift in the proportions of positive and negative cases (97% vs. 70% negative cases) during probability aggregation (FIG. 5). While slice-level metrics represented a better evaluation of the model performances, scan-level classification metrics, such as Fl scores, were deemed more appropriate when deploying such a model in practice. The model’s worse scan-level performance on the VIZ dataset could be explained by the dataset composition, including additional scans with trauma-related, tumor-related, and subdural hemorrhage besides intracerebral hemorrhage. Even though the model performance was improved by including such samples in the fine-tuning, this observation might constrain generalizability outside of explicitly trained samples showing intracerebral hemorrhage. Error! Reference source not found.

[0059] While predicting HE directly from head CT images is of interest, previous work has suffered from low generalizability, missing crucial external validation.

[0024] Applying their published model to the present data resulted in low sensitivity and poor performance in predicting the positive class, suggesting that directly predicting HE remains elusive. Instead, this project primarily focused on identifying the presence of the BHS, a visual indicator reliably associated with HE. Multi-center international data sources were harmonized for both pre-training and fine-tuning. The model demonstrated robust performance on external data, suggesting it is generalizable, excluding non-ICH images, as discussed for the VIZ dataset. Supplementation of labeled data from external sources during fine-tuning highlights improved model performance. This highlighted the potential for deploying this model across a wide range of data sources, thereby emphasizing its versatility and utility in practical applications. Given the present model results, demonstrating tradeoffs between recall and precision, the model missed up to one -third of patients with BHS while also misclassifying one in three patients as BHS positive. Therefore, the model might notNU2025-046-02 be accurate enough to be used as a single screening test to predict hematoma expansion. However, it could help identify patients requiring screening for this potentially critical finding. Future research could include training for various other indicators of HE, such as fluid levels and fully automated end-to-end HE predictions.

[0060] This study is subject to several limitations. Firstly, due to computational constraints, larger batch sizes were unable to be employed, which could have potentially enhanced the effectiveness of the pre-training phase. Additionally, while efforts were made to integrate data from various sources, the majority of the pre-training data originated from our in-house database, thus limiting the diversity of the pre-training context. Consequently, the performance and robustness of the model could be further improved by incorporating a broader range of images during acquisition. A proportion of patient demographics, such as race and ethnicity, were unavailable for ITALY and VIZ datasets. This might affect the understanding of the cohort, potentially leading to biases due to the underrepresentation of specific patient groups. Nevertheless, given the minimal differences in ICH across racial groups, any impact on model performances is expected to be minor. The prevalence of the black hole sign in the cohort was higher than in some unselected cohorts, potentially leading to altered model performance. However, the sampling was in line with the cohorts of patients who presented within the first few hours of symptom onset of intracerebral hemorrhage. One would expect the algorithm to be less useful in patients who present more than twelve hours after symptom onset when the black hole sign and subsequent hematoma expansion are uncommon. Further, the majority of patients exhibited normal international normalized ratio (INR) levels, suggesting future investigations focusing on patients with coagulopathy or undergoing anticoagulant therapy would be informative. Another potential technical limitation is that during SimCLR pre-training, adjacent correlated CT slices were treated as negative pairs for contrastive learning. Although SimCLR does not impose strict assumptions about the variability of input images and augmentations inject additional variation, the introduced contrastive noise could be mitigated by advanced sampling or loss calculation in future work. Lastly, 3D CT scans were split into 2D slices to leverage powerful 2D model architectures and training methods before aggregating the individual predictions. While we demonstrated feasibility of our 2D approach, building models based on holistic 3D scan data could provide additional value for BHS detection but comes with increased computationalNU2025-046-02 and methodological complexity. Exploring BHS prediction based on 3D data will be part of the future directions.

[0061] In summary, this study presented an innovative framework for BHS prediction, leveraging SSL on head CT images. SSL outperformed conventional supervised learning frameworks for medical imaging with consistently robust performance, indicating generalizability across internal and external validation datasets. By harnessing the semantic information in the images through contrastive self-supervised learning, the model demonstrated the potential to reduce labor, time, and costs needed for extensive data annotation. This work not only advances the field of medical image analysis in clinical neuroscience but also opens doors to broader applications of SSL in cerebrovascular diseases, where medical image encoders for various purposes could be trained using a wider range of datasets. While limitations assign the current model supporting capabilities in clinical practice to identify patients requiring further screening, continued exploration and refinement of the developed framework holds promise for further improving diagnostic accuracy and patient care.

[0062] FIG. 13 is a block diagram of an illustrative processor platform (e.g., a processor platform 1300) structured to execute the instructions of at least FIGS. 2 and 3 to implement the illustrative components, operations, and / or actions to generate information as disclosed and described herein with respect to FIGS. 1-12. The processor platform 1300 can be, for example, a server, a personal computer, a mobile device (e.g., a cell phone, a smart phone, a tablet such as an iPad™), a personal digital assistant (PDA), an Internet appliance, and / or any other type of computing device or combination of computing devices.

[0063] The processor platform 1300 includes a processor 1312. The processor 1312 is hardware. For example, the processor 1312 can be implemented by integrated circuits, logic circuits, microprocessors or controllers from any desired family or manufacturer.

[0064] The processor 1312 includes a local memory 1313 (e.g., a cache). The processor 1312 of FIG. 13 executes the instructions associated with functionality of at least FIG. 3 to implement the systems and infrastructure and associated methods of FIGS. 1-12 such as one or more computing devices of the illustrative SSL-based black hole sign detection system 300, etc. The processor 1312 is in communication with a main memory including a volatile memory 1314 and a non-volatile memory 1316 via a busNU2025-046-02 1318. The volatile memory 1314 can be implemented by Synchronous Dynamic Random-Access Memory (SDRAM), Dynamic Random-Access Memory (DRAM), RAMBUS Dynamic Random-Access Memory (RDRAM) and / or any other type of random-access memory device. The non-volatile memory 1316 can be implemented by flash memory and / or any other desired type of memory device. Access to the one or more main memory components (e.g., the volatile memory 1314, the non-volatile memory 1316) may be controlled by a clock controller.

[0065] The processor platform 1300 of the illustrated example also includes an interface circuit 1320. The interface circuit 1320 can be implemented by any type of interface standard, such as an Ethernet interface, a universal serial bus (USB), and / or a PCI express interface.

[0066] In the illustrated example, one or more input devices 1322 are connected to the interface circuit 1320. The one or more input devices 1322 permits a user to enter data and commands into the processor 1312. The one or more input devices 1322 can be implemented by, for example, a sensor, a microphone, a camera (still or video), a keyboard, a button, a mouse, a touchscreen, a trackpad, a trackball, isopoint, and / or a voice recognition system.

[0067] One or more output devices 1324 are also connected to the interface circuit 1320. The output devices 1324 can be implemented, for example, by display devices (e.g., a light emitting diode (LED), an organic light emitting diode (OLED), a liquid crystal display, a cathode ray tube display (CRT), a touchscreen, a tactile output device, and / or speakers). The interface circuit 1320, thus, typically includes a graphics driver card, a graphics driver chip, or a graphics driver processor.

[0068] The interface circuit 1320 also includes a communication device such as a transmitter, a receiver, a transceiver, a modem and / or network interface card to facilitate exchange of data with external machines (e.g., computing devices of any kind) via a network 1326 (e.g., an Ethernet connection, a digital subscriber line (DSL), a telephone line, coaxial cable, a cellular telephone system, etc.).

[0069] The processor platform 1300 also includes one or more mass storage devices 1328 for storing software and / or data. Examples of mass storage devices 1328 include floppy disk drives, hard drive disks, compact disk drives, Blu-ray disk drives, RAID systems, and digital versatile disk (DVD) drives, and / or other mass storage devices.NU2025-046-02

[0070] The coded instructions 1332 of FIG. 13 can be stored in the mass storage device 1328, in the volatile memory 1314, in the non-volatile memory 1316, and / or on a removable tangible computer readable storage medium such as a CD or DVD.

[0071] According to an aspect of the present disclosure, a computer-implemented method for prediction of hematoma expansion is provided. The method includes retrieving, from a data store, a dataset comprising a set of computed tomography (CT) scans for a population of individuals suffering a stroke. The method includes annotating the dataset, wherein the annotating comprises scan-level annotation and slice-level annotation. The method includes splitting the dataset into a pre-training dataset and a fine-tuning dataset. The method includes training, based on the pre -training dataset and the fine-tuning dataset, a self-supervised learning (SSL) model in a pre -training phase and a fine-tuning phase. The method includes finetuning a downstream classifier to execute slice-level black hole sign identification. The method includes processing, by the SSL model, a diagnostic scan of a patient, wherein the diagnostic scan of the patient comprises a plurality of slices. The method includes generating, based on processing of the diagnostic scan of the patient, probabilities for each slice of the plurality of slices of the diagnostic scan, wherein the probabilities are associated with identification of a black hole sign. The method includes predicting, based on an aggregation of the probabilities for each slice of the plurality of slices of the diagnostic scan, a probability of hematoma expansion for the patient.

[0072] According to other aspects of the present disclosure, the computer-implemented method may include one or more of the following features. The annotating may comprise annotating the dataset with information corresponding to indications of individuals suffering a stroke. The set of CT scans may comprise patterns of hemorrhage characteristic of trauma, tumors, subdural hematoma, and hemorrhages consistent with spontaneous intracerebral hemorrhage. The dataset may comprise over one million scans. The set of CT scans may comprise axial head CT scans. Each CT scan of the set of CT scans may be clipped to a brain window. The method may further comprise reslicing each set of CT scans to obtain equidistant slices. A standard slice thickness may comprise 3 mm for infratentorial brain regions and 5 mm for supratentorial brain regions. The pre -training dataset and the fine-tuning dataset may be split to ensure slices associated with particular patients appear in one of the pre-training dataset or the fine-tuning dataset. The model may comprise a ResNet50 encoder within a simpleNU2025-046-02 framework for contrastive learning of visual representation (SimCLR) as a common framework to implement self-supervised learning.

[0073] According to another aspect of the present disclosure, a computing platform for prediction of hematoma expansion is provided. The computing platform includes a processor. The computing platform includes a memory in communication with the processor and storing instructions that, when executed by the processor, cause the computing platform to retrieve, from a data store, a dataset comprising a set of computed tomography (CT) scans for a population of individuals suffering a stroke. The instructions cause the computing platform to annotate the dataset, wherein the annotating comprises scan-level annotation and slice-level annotation. The instructions cause the computing platform to split the dataset into a pre-training dataset and a fine- tuning dataset. The instructions cause the computing platform to train, based on the pre-training dataset and the fine-tuning dataset, a self-supervised learning (SSL) model in a pre-training phase and a fine-tuning phase. The instructions cause the computing platform to finetune a downstream classifier to execute slice-level black hole sign identification. The instructions cause the computing platform to process, by the SSL model, a diagnostic scan of a patient, wherein the diagnostic scan of the patient comprises a plurality of slices. The instructions cause the computing platform to generate, based on processing of the diagnostic scan of the patient, probabilities for each slice of the plurality of slices of the diagnostic scan, wherein the probabilities are associated with identification of a black hole sign. The instructions cause the computing platform to predict, based on an aggregation of the probabilities for each slice of the plurality of slices of the diagnostic scan, a probability of hematoma expansion for the patient.

[0074] According to other aspects of the present disclosure, the computing platform may include one or more of the following features. The annotating may comprise annotating the dataset with information corresponding to indications of individuals suffering a stroke. The set of CT scans may comprise patterns of hemorrhage characteristic of trauma, tumors, subdural hematoma, and hemorrhages consistent with spontaneous intracerebral hemorrhage. The dataset may comprise over one million scans. The set of CT scans may comprise axial head CT scans. Each CT scan of the set of CT scans may be clipped to a brain window. The instructions may further cause the computingNU2025-046-02 platform to re-slice each set of CT scans to obtain equidistant slices. A standard slice thickness may comprise 3 mm for infratentorial brain regions and 5 mm for supratentorial brain regions. The pre-training dataset and the fine-tuning dataset may be split to ensure slices associated with particular patients appear in one of the pretraining dataset or the fine-tuning dataset. The model may comprise a ResNet50 encoder within a simple framework for contrastive learning of visual representation (SimCLR) as a common framework to implement self-supervised learning.Funding

[0075] This work was supported in part by NIH grant R01LM013337 and R01NS110779. AZ reports speaker honoraria from CSL Behring, Alexion-Astra Zeneca and Daiichi Sankyo and advisory board honoraria from Bayer, Astra Zeneca, all outside of the submitted work.Data availability

[0076] The data that support the findings of this study are available from the corresponding author, upon reasonable request.References

[0077] 1. Virani SS, Alonso A, Benjamin EJ, et al. Heart disease and stroke statistics — 2020 update: a report from the American Heart Association. Circulation 2020; 141(9): el39- e596.

[0078] 2. Qureshi Al, Mendelow AD, Hanley DF. Intracerebral haemorrhage. The Lancet 2009; 373(9675): 1632-44.

[0079] 3. Pinho J, Costa AS, Araujo JM, Amorim JM, Ferreira C. Intracerebral hemorrhage outcome: a comprehensive update. Journal of the neurological sciences 2019; 398: 54- 66.

[0080] 4. Vespa PM, O’phelan K, Shah M, et al. Acute seizures after intracerebral hemorrhage: a factor in progressive midline shift and outcome. Neurology 2003; 60(9): 1441-6.

[0081] 5. Broderick JP, Brott TG, Duldner JE, Tomsick T, Huster G. Volume of intracerebral hemorrhage. A powerful and easy-to-use predictor of 30-day mortality. Stroke 1993; 24(7): 987-93.NU2025-046-02

[0082] 6. Marushi M. Shima T, Okada Y, Nishida M, Yamane K. Involvement of fluctuating high blood pressure in the enlargement of spontaneous intracerebral hematoma. Neurologia medico-chirurgica 2001; 41(6): 300-5.

[0083] 7. Ohwaki K, Yano E, Nagashima H, Hirata M, Nakagomi T, Tamura A. Blood pressure management in acute intracerebral hemorrhage: relationship between elevated blood pressure and hematoma enlargement. Stroke 2004; 35(6): 1364-7.

[0084] 8. Rodriguez-Luna D, Rodriguez-Villatoro N, Juega JM, et al. Prehospital systolic blood pressure is related to intracerebral hemorrhage volume on admission. Stroke 2018; 49(1): 204-6.

[0085] 9. Rodriguez-Luna D, Pineiro S, Rubiera M, et al. Impact of blood pressure changes and course on hematoma growth in acute intracerebral hemorrhage. European journal of neurology 2013; 20(9): 1277-83.

[0086] 10. Anderson CS, Huang Y, Wang IG, et al. Intensive blood pressure reduction in acute cerebral haemorrhage trial (INTERACT): a randomised pilot trial. The Lancet Neurology 2008; 7(5): 391-9.

[0087] 11. Blacquiere D, Demchuk AM, ALHazzaa M, et al. Intracerebral hematoma morphologic appearance on noncontrast computed tomography predicts significant hematoma expansion. Stroke 2015; 46(11): 3111-6.

[0088] 12. Li Q, Liu Q-J, Yang W-S, et al. Island sign: an imaging predictor for early hematoma expansion and poor outcome in patients with intracerebral hemorrhage. Stroke 2017; 48(11): 3019-25.

[0089] 13. Li Q, Zhang G, Huang Y-J, et al. Blend sign on computed tomography: novel and reliable predictor for early hematoma growth in patients with intracerebral hemorrhage. Stroke 2015; 46(8): 2119-23.

[0090] 14. Ng D, Churilov L, Mitchell P, Dowling R, Yan B. The CT swirl sign is associated with hematoma expansion in intracerebral hemorrhage. American loumal of Neuroradiology 2018; 39(2): 232-7.

[0091] 15. Demchuk AM, Dowlatshahi D, Rodriguez-Luna D, et al. Prediction of haematoma growth and outcome in patients with intracerebral haemorrhage using the CT- angiography spot sign (PREDICT): a prospective observational study. The Lancet Neurology 2012; 11(4): 307-14.NU2025-046-02

[0092] 16. Parikh NS, Kamel H, Navi BB, et al. Liver fibrosis indices and outcomes after primary intracerebral hemorrhage. Stroke 2020; 51(3): 830-7.

[0093] 17. Zhang F, Li H, Qian J, et al. Hyperglycemia predicts blend sign in patients with intracerebral hemorrhage. Medical Science Monitor: International Medical Journal of Experimental and Clinical Research 2018; 24: 6237.

[0094] 18. Zhang F, Zhang S, Tao C, et al. Association between serum glucose level and spot sign in intracerebral hemorrhage. Medicine 2019; 98(11).

[0095] 19. Silva Y, Leira R, Tejada J, et al. Molecular signatures of vascular injury are associated with early growth of intracerebral hemorrhage. Stroke 2005; 36(1): 86-91.

[0096] 20. Di Napoli M, Parry-Jones AR, Smith CJ, et al. C-reactive protein predicts hematoma growth in intracerebral hemorrhage. Stroke 2014; 45(1): 59-65.

[0097] 21. Tanioka S, Yago T, Tanaka K, et al. Machine learning prediction of hematoma expansion in acute intracerebral hemorrhage. Scientific Reports 2022; 12(1): 12452.

[0098] 22. Liu J, Xu H, Chen Q, et al. Prediction of hematoma expansion in spontaneous intracerebral hemorrhage using support vector machine. EBioMedicine 2019; 43: 454- 9.

[0099] 23. Ma C, Zhang Y, Niyazi T, et al. Radiomics for predicting hematoma expansion in patients with hypertensive intraparenchymal hematomas. European journal of radiology 2019; 115: 10-5.

[0100] 24. Tran AT, Zeevi T, Haider SP, et al. Uncertainty-aware deep-learning model for prediction of supratentorial hematoma expansion from admission non-contrast head computed tomography scan. NPJ Digital Medicine 2024; 7(1): 26.

[0101] 25. Li Q, Yang W-S, Chen S-L, et al. Black hole sign predicts poor outcome in patients with intracerebral hemorrhage. Cerebrovascular Diseases 2018; 45(1-2): 48-53.

[0102] 26. Li Q, Zhang G, Xiong X, et al. Black hole sign: novel imaging marker that predicts hematoma growth in patients with intracerebral hemorrhage. Stroke 2016; 47(7): 1777- 81.

[0103] 27. Huang S-C, Pareek A, Jensen M, Lungren MP, Yeung S, Chaudhari AS. Selfsupervised learning for medical image classification: a systematic review and implementation guidelines. NPJ Digital Medicine 2023; 6(1): 74.NU2025-046-02

[0104] 28. Morotti A, Boulouis G, Dowlatshahi D, et al. Standards for detecting, interpreting, and reporting noncontrast computed tomographic markers of intracerebral hemorrhage expansion. Annals of neurology 2019; 86(4): 480-92.

[0105] 29. ARELIUS. CT Head Scans (jpg files). 2020. https: / / www.kaggle.com / datasets / clarksaben / ct-head-scans (accessed Sep 29, 2023).

[0106] 30. Chilamkurthy S, Ghosh R, Tanamala S, et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. The Lancet 2018; 392(10162): 2388-96.

[0107] 31. Hssayeni M, Croock M, Salman A, ALkhafaji H, Yahya Z, Ghoraani B. Computed tomography images for intracranial hemorrhage detection and segmentation. Intracranial Hemorrhage Segmentation Using A Deep Convolutional Model Data 2020; 5(1): 14.

[0108] 32. Kitamura FC. Head CT - hemorrhage. In: Kaggle, editor.; 2018.

[0109] 33. Chen T, Kornblith S, Norouzi M, Hinton G. A simple framework for contrastive learning of visual representations. International conference on machine learning; 2020: PMLR; 2020. p. 1597-607.

[0110] 34. Chattopadhay A, Sarkar A, Howlader P, Balasubramanian VN. Grad-cam++:Generalized gradient-based visual explanations for deep convolutional networks. 2018 IEEE winter conference on applications of computer vision (WACV); 2018: IEEE; 2018. p. 839-47.

[0111] 35. Tejani AS, Klontzas ME, Gatti AA, et al. Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 Update. Radiology: Artificial Intelligence 2024: e240300.

[0112] One or more aspects discussed herein can be embodied in computer-usable or readable data and / or computer-executable instructions, such as in one or more program modules, executed by one or more computers or other devices as described herein. Generally, program modules include routines, programs, objects, components, data structures, and the like, that perform particular tasks or implement particular abstract data types when executed by a processor in a computer or other device. The modules can be written in a source code programming language that is subsequently compiled for execution, or can be written in a scripting language such as (but not limited to) HTML or XML. The computer executable instructions can be stored on a computer readable medium suchNU2025-046-02 as a hard disk, optical disk, removable storage media, solid-state memory, RAM, and the like. As will be appreciated by one of skill in the art, the functionality of the program modules can be combined or distributed as desired in various embodiments. In addition, the functionality can be embodied in whole or in part in firmware or hardware equivalents such as integrated circuits, field programmable gate arrays (FPGA), and the like. Particular data structures can be used to implement one or more aspects more effectively and are discussed herein, and such data structures are contemplated within the scope of computer executable instructions and computer- usable data described herein. Various aspects discussed herein can be embodied as a method, a computing device, a system, and / or a computer program product.

[0113] Although the present invention has been described in certain specific aspects, many additional modifications and variations would be apparent to those skilled in the art. In particular, any of the various processes described above can be performed in alternative sequences and / or in parallel (on different computing devices) to achieve similar results in a manner that is more appropriate to the requirements of a specific application. It is therefore to be understood that the present invention can be practiced otherwise than specifically described without departing from the scope and spirit of the present invention. Thus, embodiments of the present invention should be considered in all respects as illustrative and not restrictive.

Claims

NU2025-046-02 WHAT IS CLAIMED IS:

1. A computer-implemented method for prediction of hematoma expansion, comprising: retrieving, from a data store, a dataset comprising a set of computed tomography (CT) scans for a population of individuals suffering a stroke;annotating the dataset, wherein the annotating comprises scan-level annotation and slice-level annotation;splitting the dataset into a pre-training dataset and a fine-tuning dataset; training, based on the pre -training dataset and the fine-tuning dataset, a self-supervised learning (SSL) model in a pre-training phase and a fine-tuning phase;finetuning a downstream classifier to execute slice-level black hole sign identification; processing, by the SSL model, a diagnostic scan of a patient, wherein the diagnostic scan of the patient comprises a plurality of slices;generating, based on processing of the diagnostic scan of the patient, probabilities for each slice of the plurality of slices of the diagnostic scan, wherein the probabilities are associated with identification of a black hole sign; andpredicting, based on an aggregation of the probabilities for each slice of the plurality of slices of the diagnostic scan, a probability of hematoma expansion for the patient.

2. The computer-implemented method of claim 1, wherein the annotating comprises annotating the dataset with information corresponding to indications of individuals suffering a stroke.

3. The computer-implemented method of claim 1, wherein the set of CT scans comprise patterns of hemorrhage characteristic of trauma, tumors, subdural hematoma, and hemorrhages consistent with spontaneous intracerebral hemorrhage.

4. The computer-implemented method of claim 1, wherein the dataset comprises over one million scans.

5. The computer-implemented method of claim 1, wherein the set of CT scans comprise axial head CT scans.NU2025-046-02 6. The computer-implemented method of claim 1, wherein each CT scan of the set of CT scans is clipped to a brain window.

7. The computer-implemented method of claim 1, further comprising re-slicing the set of CT scans to obtain equidistant slices.

8. The computer-implemented method of claim 7, wherein a standard slice thickness comprises 3 mm for infratentorial brain regions and 5 mm for supratentorial brain regions.

9. The computer-implemented method of claim 1, wherein the pre-training dataset and the fine-tuning dataset are split to ensure slices associated with particular patients exclusively appear in one of the pre-training dataset or the fine-tuning dataset.

10. The computer-implemented method of claim 1, wherein the SSL model comprises a ResNet50 encoder within a simple framework for contrastive learning of visual representation (SimCLR) as a common framework to implement self-supervised learning.

11. A computing platform for prediction of hematoma expansion, comprising:a processor; anda memory in communication with the processor and storing instructions that, when executed by the processor, cause the computing platform to:retrieve, from a data store, a dataset comprising a set of computed tomography (CT) scans for a population of individuals suffering a stroke;annotate the dataset with scan-level annotation and slice-level annotation; split the dataset into a pre-training dataset and a fine-tuning dataset; train, based on the pre -training dataset and the fine-tuning dataset, a selfsupervised learning (SSL) model in a pre-training phase and a fine-tuning phase; finetune a downstream classifier to execute slice-level black hole sign identification;process, by the SSL model, a diagnostic scan of a patient, wherein the diagnostic scan of the patient comprises a plurality of slices;generate, based on processing of the diagnostic scan of the patient, probabilities for each slice of the plurality of slices of the diagnostic scan, wherein the probabilities are associated with identification of a black hole sign; andNU2025-046-02 predict, based on an aggregation of the probabilities for each slice of the plurality of slices of the diagnostic scan, a probability of hematoma expansion for the patient.

12. The computing platform of claim 11 , wherein the the dataset is annotated with information corresponding to indications of individuals suffering a stroke.

13. The computing platform of claim 11, wherein the set of CT scans comprise patterns of hemorrhage characteristic of trauma, tumors, subdural hematoma, and hemorrhages consistent with spontaneous intracerebral hemorrhage.

14. The computing platform of claim 11 , wherein the dataset comprises over one million scans.

15. The computing platform of claim 11, wherein the set of CT scans comprise axial head CT scans.

16. The computing platform of claim 11, wherein each CT scan of the set of CT scans is clipped to a brain window.

17. The computing platform of claim 11, further comprising re-slicing each the of CT scans to obtain equidistant slices.

18. The computing platform of claim 17, wherein a standard slice thickness comprises 3 mm for infratentorial brain regions and 5 mm for supratentorial brain regions.

19. The computing platform of claim 11, wherein the pre-training dataset and the fine-tuning dataset are split to ensure slices associated with particular patients exclusively appear in one of the pre-training dataset or the fine-tuning dataset.

20. The computing platform of claim 11, wherein the SSL model comprises a ResNet50 encoder within a simple framework for contrastive learning of visual representation (SimCLR) as a common framework to implement self-supervised learning.

Images