Background & Summary

Glioblastoma (GBM) is the most prevalent and aggressive primary brain tumor, with a dismal prognosis and roughly 12 000 new cases diagnosed each year1. Despite decades of research, overall survival for high-grade gliomas (WHO grades III–IV) has changed little. Major challenges include marked inter- and intratumoral heterogeneity, microscopic invasion and the difficulty of delineating tumor margins during surgery2. GBM is typified by profound intratumoural heterogeneity and immune evasion, driven by oncogenic processes such as aberrant growth-factor signalling, dysregulated cell-cycle control, pathological angiogenesis, invasive migration, genomic instability and defects in apoptotic pathways. Histopathologically, it shows diffuse astrocytic infiltration, pseudopalisading necrosis, extensive microvascular proliferation and high mitotic activity3,4.

For diagnosing GBM, a range of advanced clinical imaging techniques is utilized5. For example, magnetic resonance imaging (MRI), often enhanced with contrast agents, improves diagnostic clarity. Computed tomography (CT) and positron emission tomography (PET) scans also provide information about tumor size and location5,6. The latest guidelines from the European Association for Neuro-Oncology (EANO) and the Society for Neuro-Oncology (SNO) recommend a multimodal standard-of-care: maximal safe resection followed by concurrent radiotherapy and temozolomide, then adjuvant temozolomide. Molecular profiling (e.g., MGMT promoter methylation, IDH mutation status) is used to guide prognosis and determine clinical trial eligibility. For recurrent disease, selective re-irradiation is advised when feasible. Additionally, emerging modalities—such as tumor-treating fields, immune checkpoint inhibitors, personalized vaccines, and oncolytic viral therapies—are increasingly incorporated through clinical trials7,8. Despite aggressive treatment, GBM often recurs due to microscopic residual tumor cells that remain in the surrounding brain tissue. Enhanced visualization of these residual cells could facilitate more precise interventions and monitoring, potentially reducing recurrence and improving patient outcomes9.

Ultrasound is another modality frequently used to deliver real-time, non-invasive imaging of soft tissues and to monitor therapeutic outcomes10. During GBM preclinical studies, ultrasound imaging can help distinguish tumor tissue from healthy brain tissue by detecting differences in acoustic properties. These variations produce distinct signal patterns, enabling real-time visualization of tumor margins for more accurate diagnosis and treatment planning. Meta-analyses show that intraoperative ultrasound helps reduce residual tumor in GBM surgery by clearly visualizing tumor margins. Its real-time imaging supports more informed decisions and improves surgical precision11. In addition to its imaging capabilities, ultrasound is being explored for therapeutic uses, especially in non-invasive or minimally invasive treatments like focused ultrasound for targeted drug delivery, disrupting the blood-brain barrier, and tumor ablation12,13,14. Among the various therapeutic ultrasound methods, one promising approach is histotripsy, which has gained attention for its potential to target and ablate tumors precisely15.

Histotripsy is a non-thermal ultrasound technique that uses short, focused high-pressure pulses to destroy tumor tissue through cavitation bubbles. It shows promise as a minimally invasive alternative to surgery and other ablation methods16. Ongoing pre-clinical studies have also demonstrated that histotripsy can trigger a strong immune response by releasing tumor antigens during the destruction of cancer cells. These antigens stimulate the immune system, helping it recognize and attack any residual tumor cells. This immune activation could also enhance the overall effectiveness of the treatment by reducing the likelihood of tumor recurrences17.

The GL261 glioma cell line is commonly used in glioblastoma research for its ability to mimic the aggressive nature of human GBM. It enables studies on tumor growth and treatment response. Histotripsy has been applied to GL261 tumors in mice, showing promising potential for advancing cancer therapy18,19. In these studies, precise tumor delineation is essential for effective histotripsy. By clearly defining the tumor boundaries, histotripsy can accurately target the cancerous tissue, sparing surrounding healthy areas20. GL261 tumor segmentation has conventionally been performed manually using MRI scans, where tumor boundaries are identified and delineated directly from the imaging data21. Machine learning (ML) refers to a subset of artificial intelligence (AI) that uses algorithms and statistical models to identify patterns and make predictions based on data. Automatic detection, segmentation, and classification of tumors are among important applications of ML models in clinical settings22,23. These models can handle different imaging modalities, including MRI, CT, and ultrasound.

The number of studies utilizing ML models for automated tumor delineation has steadily increased, with many demonstrating substantial improvements in both accuracy and efficiency over manual methods. For example, Weng et al.24 conducted a multicenter study to assess the feasibility and effectiveness of AI-based automatic segmentation of rodent brain tumors on MRI. Utilizing a 3D U-Net architecture, the study achieved accurate segmentation of brain tumors in MRI images from 57 WAG/Rij rats and 46 mice. Lee et al.25 developed TumorPrism3D, a software tool designed for AI-assisted segmentation of brain tumors in MR images. The study tested the software on MRI data from 185 GBM patients, focusing on segmenting contrast-enhancing lesions, necrotic regions, and non-enhancing T2 high signal intensity components. TumorPrism3D demonstrated high accuracy in segmenting these tumor components, proving effective in facilitating the analysis of brain tumors in MRI scans. Di Ieva et al.26 conducted a study to assess the effectiveness of a machine learning model for the automatic segmentation of brain tumors on MRI, particularly for clinical applications such as radiosurgery planning. The model was trained on the Medical Decathlon dataset and tested on the BraTS 2019 glioma dataset. The study demonstrated that the model accurately segmented brain tumors using various MRI sequences and could generalize well to non-glioma cases through transfer learning.

Our group’s overarching goal is to use histotripsy to ablate the entire volume of a GL261-transplanted mouse brain tumor using a device under automated robotic arm control. This approach aims to enhance ablation accuracy and precision while reducing procedure time. To achieve this, we first need to delineate the GBM boundary. In pursuit of this larger goal, we began by collecting the first ultrasound datasets of mice with GL261 tumors using a custom-built imaging system27 capable of performing both histotripsy and high-resolution ultrasound imaging. After collecting ultrasound images of the brains of ten in-vivo mice (5 tumor-bearing and 5 non-tumorous) and two ex-vivo mice (tumor-bearing), the dataset was annotated by a group of five trained students and verified by two experienced professionals in the field. The Dice score, used for segmentation evaluation, measures the overlap between two data sets, which is commonly applied in image segmentation problems. A higher Dice score indicates a greater degree of overlap, thus reflecting the quality and accuracy of the segmentations. This metric ensures the reliability of the dataset and confirms the consistency of the annotations between both groups. Also, to ensure the highest quality of annotation, the corresponding MRI and histology images, when available, were also considered during the annotation process.

The dataset is intended to train machine learning models to accurately segment tumor boundaries—a task that remains challenging during surgery. A total of 1,856 ultrasound images were collected, including 1,448 tumor images and 408 non-tumor images. This study aims to enhance surgical outcomes and reduce the time and costs associated with tumor boundary identification, thereby laying the groundwork for improved GBM surgeries in humans. Additionally, the dataset serves as a bridge between preclinical research and clinical application, supporting more precise tumor resections in both experimental models and future clinical practice.

Methods

Eleven female and one male C57BL/6 mice (~6 weeks old, weight range 22–25.5 g) were purchased from Charles River Laboratories and housed in clean, ventilated cages. The study, conducted with ethical approval from the Dalhousie University Committee on Laboratory Animals (protocol no. 24-012), adhered to the guidelines of the Canadian Council on Animal Care. Ultrasound data were collected during both ex-vivo and in-vivo experiments, ensuring a comprehensive dataset. The annotation was performed collaboratively by five students using custom-developed software, under the guidance of two experienced professionals who ensured accuracy and consistency. As illustrated in Fig. 1, the study’s workflow was structured around four key steps: mouse preparation, data collection, data annotation, and data evaluation. This methodical approach allowed for efficient and precise analysis of the ultrasound data, leading to high-quality results.

Fig. 1
figure 1

(a) Twelve mice were injected with GL261 glioma cells and (b) monitored via MRI every five days. Tumors typically became detectable between days 7 and 14, with imaging continued until tumor size reached ~2 mm— (c) marking the transition point to ultrasound imaging. (d) Ultrasound videos from both in vivo and ex vivo experiments were analyzed through a collaborative process involving students and clinical experts. Five independent annotators segmented the data, (e) and their outputs were merged and refined by an expert to generate the final ground truth. (f) Whole mouse brain extracted postmortem for histological processing and (g) ground truth segmentations were validated using MRI and histological analysis following brain extraction.

Full size image

Mouse preparation

Ten mice (from twelve) underwent standardized tumor induction procedures which were described in detail previously27. Each mouse received the first dose of meloxicam (5 mg/kg, subcutaneously) 30 to 40 minutes before surgery to control pain. Then, the mice were anesthetized with isoflurane (5% induction, ~2% maintenance), and a small borehole was created in the skull to target one of three different cell injection sites in the brain (see Table 1) using the coordinate notation system of a mouse brain atlas28. In stereotaxic surgery based on a standardized brain atlas the coordinates AP, ML, and DV refer to anterior-posterior, medial-lateral, and dorsal-ventral positions, respectively. The AP and DV axes are measured relative to the skull reference point, Bregma, with negative AP values indicating a posterior position (toward the back) and positive values indicating an anterior position (toward the front). The ML axis measures the lateral distance from the midline of the brain, where positive values represent positions to the right and negative values to the left. The DV axis represents the vertical depth, with negative values indicating positions deeper within the brain tissue. These coordinates are essential for achieving the high precision required for accurate injections or surgical interventions.

Table 1 Overview of mouse cohort and injection sites for GL261 tumor induction.
Full size table

After drilling the borehole, the needle was lowered to the target DV coordinate and 2 µl of GL261 cells (≈50,000 cells) was injected at 0.1 µl/min. The incision was then sutured and lidocaine applied for local pain relief. Post-op, mice received mash, recovery gel, and water and were watched until fully mobile. At 24 h, they were given meloxicam in 0.5 ml saline subcutaneously, underwent a clinical exam and weight check, and had daily cage-side observations. Analgesics, fluids, gel, and mash were provided as needed thereafter. Weekly clinical exams continued until study endpoints; incisions were checked daily and sutures removed on days 7–10.

Starting from the fifth-day post-operation, the mice were anesthetized with isoflurane, and tumor growth was measured every five days by MRI scan. Imaging was performed using a 3 T MRI system with a 21 cm inner diameter gradient coil and a Varian direct drive console. The RF coil used was a 30 mm quadrature transmit/receive coil, optimized for small animal imaging. The T2-weighted fast spin-echo (FSE) protocol employed a 64 × 64 matrix, a 17 × 17 mm field of view, and 20 slices of 0.7 mm thickness. When the tumor reached 2–3 mm, mice were prepared for terminal procedures within two days. Under isoflurane anesthesia, the head was secured in a stereotaxic frame, a scalp incision was made, and an 8 mm craniectomy centered on the injection site exposed the brain and dura. Ultrasound gel was applied, and tumor imaging was performed.

For image collection, the probe was positioned above the tumor site, imaging through the craniectomy opening (details of the imaging system are given below). Ultrasound imaging was conducted from multiple angles and across various brain slices to capture a broad range of perspectives. In addition to the tumor-bearing mice, non-tumor mice were included in the study to provide control images for comparison. Histotripsy was also performed in some mice for a related study outside the scope of this article, which limited the availability of untreated tissue histology in some animals. After imaging, mice were euthanized with sodium pentobarbital (150 mg/kg i.p.) and perfused with 10% formalin. Brains were fixed overnight, soaked in 30% sucrose, embedded in OCT, and cut coronally at 15 µm. Sections were H&E stained and scanned at 20× magnification.

During histological processing, occasional artifacts affected section quality. Slight, uneven shrinkage during fixation and sucrose infiltration caused minor contractions in section dimensions, and rapid freezing in OCT sometimes introduced localized stresses that made a few slices curve or cup rather than lie flat. Mechanical shear at the cryostat blade edge could generate micro-tears or gentle folds in regions of varying tissue density, and human error during sectioning occasionally distorted or partly removed the region of interest (ROI). Although most sections remained usable, these sporadic deformations shifted anatomical landmarks and complicated precise alignment with the ultrasound and MRI datasets. Figure 2 displays a high-resolution ultrasound image alongside the corresponding histological section of the same tissue and MRI. This pairing highlight how detailed structural features captured in the ultrasound correlate with the microscopic anatomy revealed through histology and MRI imaging, especially the tumor.

Fig. 2
figure 2

(a) H&E image, and (b) MRI with corresponding (c) high-quality ultrasound image (d) brain surface and tumor outlines for each modality, showing very good agreement in tumor boundary position. A schematic of some major brain anatomical features is also shown for reference.

Full size image

To identify tumor boundaries post-surgically, MRI and H&E-stained histological sections were utilized when available. To enable a semi-quantitative assessment of spatial concordance across ultrasound, MRI, and H&E, tumor boundaries and key anatomical landmarks—such as the cortical surface, striatal boundary, and visible tumor margins—were first defined independently in each modality. The spatial scale of each dataset was then adjusted to ensure consistent units, allowing meaningful spatial comparison. Using the identified landmarks, a spatial transform was computed to align the ultrasound images with the MRI and H&E datasets. This transform was subsequently applied to the entire ultrasound image to bring its anatomy into approximate alignment with the other modalities.

Although perfect co-registration is limited by intrinsic differences in acquisition technique, image resolution, and deformation artifacts introduced during histological, close spatial correspondence was achieved in most cases. For tumor delineation in ultrasound, an averaged boundary was derived from the partially aligned contours to accommodate minor discrepancies and provide a consistent segmentation reference. Figure 2(d), black arrows indicate regions where tumor boundaries from ultrasound, MRI, and H&E are closely aligned.

Image acquisition

Ultrasound imaging was conducted using a custom-built beamformer. The detailed specifications of the hardware, construction methods, and imaging system processing techniques are extensively covered in prior publications29,30. Here, only a brief overview is provided. The system featured a 64-element phased array with an element pitch of 48 µm and an elevation length of 2 mm. The elements were defined by laser micro-machining a raw PIN-PMN-PT single crystal piezoelectric substrate. Aluminum wire bonds were used to connect the array elements to a cable interconnect PCB and then subsequently potted with an insulating epoxy. A 21 µm thick parylene-C acoustic matching layer was vacuum deposited over the array and then a polyurethane lens was cast for focusing the elevation dimension to a 7 mm depth. The center frequency of the pulse was 30 MHz, and the probe could easily be co-aligned with a therapeutic histotripsy device with a small hole in the center. The FPGA-based imaging system provided clear images with axial and lateral resolutions of 40 µm and 130 µm, respectively, covering a 2–15 mm depth range, 128-line angles, and a 64° sector range. Processed image data was displayed using the Sagacity software platform, developed in collaboration with Daxsonics Ultrasound Inc. (Halifax, Canada), which facilitated both raw data acquisition and proprietary enhanced image processing.

Although the beamformer generated high-resolution images with minimal electronic noise, occasional artifacts were observed. Speckle noise, resulting from scattering by microscopic tissue structures, sometimes reduced contrast at the tumor margins, making accurate delineation more difficult in certain frames. Hemorrhagic signals, appearing as hyperechoic streaks or patches due to the presence of extravasated blood cells or small gas bubbles, occasionally resembled solid tissue and may obscure the true tumor edge. Other common artifacts were grating lobes, where bright structures such as bone, even when out of frame, can cause blurry streaks to appear on the opposite side of the image; and reverberation artifacts, where ultrasound waves reflected multiple times between the probe and a structure, producing faint duplicate images at greater depths.

Image annotation

The annotation team comprised two experienced professionals and five students. Prior to initiating annotation tasks, the team underwent thorough training, which included in-depth familiarization with the GL261 mouse tumor models and key technical aspects of ultrasound imaging for accurate data interpretation. This training was delivered through a combination of online sessions and in-person guidance by the experts. During the annotation training process, any annotations that did not meet the required standards were returned to the respective student for further refinement. Annotators used custom-developed software programmed in the Python programming language that was designed for precise pixel-wise closed shape segmentation. This software has been made publicly available on Figshare for reproducibility and broader access. A total of 1,856 ultrasound images were collected from a mouse tumor injected with GL261, including 1,448 tumor images and 408 non-tumor images. The final segmentation was represented in a binary format, with tumor regions depicted in white and non-tumor areas in black. During the official annotation phase, each image was annotated by five annotators, and any overlapping pixels were reviewed for consistency.

Data Records

All data records are available on Figshare31. The unzipped dataset folder contains the original ultrasound images, corresponding masks, and both raw and enhanced video recordings. Additionally, an Excel file (“Read Me”) is included, providing detailed information about the mouse tumor, mouse gender, imaging plane, and whether the data was collected in-vivo or ex-vivo. Also, each recording file is organised into two folders: “Images,” and “Masks”. Additionally, there is an MP4 file that contains the raw and enhanced video recording. The images in the “Images” folder correspond directly to the masks in the “Masks” folder, with matching names for easy reference. Images are named as “frame n_img,” where “n” represents the frame number. All files are provided in PNG format, with image dimensions of 780 × 900 mm/pixel, 16-bit unsigned integer bit depth, and little-endian byte order.

Technical Validation

In this study, the ultrasound images underwent multiple rounds of annotation to ensure accuracy and consistency. Initially, each image was annotated five times by different annotators. The resulting segmentations were then averaged to create a mean segmentation, which was further refined by professional experts to establish a more precise ground truth. To validate the accuracy of these annotations, corresponding histology and MRI images were used where available, ensuring that the segmentations aligned with the actual tumor boundaries observed in these modalities, allowing for some discrepancy due to differences in exact slice position, and in the case of comparing MRI to ultrasound and histology, the tissue distortion that occurs after opening the skull.

To assess the consistency and reliability of the annotations, a subset of 30 images was randomly selected from the full dataset for validation. Selection was stratified to ensure that the subset spanned the entire observed range of tumor diameters (2.2–5.0 mm), thereby including small, medium, and large lesions (Fig. 3). This stratified sampling allowed us to evaluate annotation performance across diverse tumor presentations and ensured that the Dice-based validation reflected the dataset’s full morphological variability. The Dice coefficient, a measure of overlap between two segmentations ranging from 0 (no overlap) to 1 (perfect overlap), was calculated by comparing the mean of these segmentations to the annotations made by two professional experts. This process allowed for the evaluation of both intra-annotator and inter-annotator consistency. The mean Dice coefficient for the 30 images, comparing the students’ mean segmentations to the expert annotations, was 88 ± 0.05 for the first professional expert and 90 ± 0.07 for the second professional expert. These results demonstrate the stability and consistency of the annotations across different annotators and time points, confirming the dataset’s reliability for further research and analysis. Figure 4 provides a visual example of this validation process.

Fig. 3
figure 3

Distribution of tumor diameters in the validation dataset used for annotation consistency analysis.

Full size image
Fig. 4
figure 4

(a) Mean segmentation derived from five independent annotations for a sample image, (b) comparison between the mean segmentation and the professional expert’s annotation, and (c) visual overlay showing the Dice score and alignment between the segmentations.

Full size image

Usage Notes

The dataset comprises ultrasound images, corresponding binary segmentation masks, and both raw and enhanced video recordings, all organized into structured, easily accessible folders. The segmentation masks are precisely aligned with the images, facilitating seamless integration into ML workflows. The accompanying “Read Me” file provides detailed information on the imaging plane and acquisition speed, ensuring clarity and ease of use for further analysis.

Additionally, the custom segmentation software developed for this dataset, available on Figshare32, can assist in segmentation tasks. After running the software, the folder includes the images selected for segmentation. If the user needs to save the segmentation in another folder, they can choose a new directory after clicking “Save in another folder.” For segmentation, they need to select the boundary of the tumor, which is displayed as a red line. After pressing “Save Mask,” the corresponding mask will be saved in the specified directory with an aligned file name.