Hospital AI platform lifts brain bleed detection accuracy in three-month real-world trial

Radiologists in Oslo have shown that an in-house artificial intelligence tool for spotting brain bleeds on head CT scans can be substantially improved when it is wired directly into a hospital’s imaging workflow and retrained on feedback from the clinicians who use it. Over a three-month period, the team behind the NeoMedSys platform reported that the VIOLA-AI model’s sensitivity rose from 79.2% to 90.3%, its specificity climbed from 80.7% to 89.3%, and the area under its receiver operating characteristic curve increased from 0.873 to 0.949, according to a study published in BMC Methods.

The study, led by researchers at Oslo University Hospital, was designed as a prospective pragmatic investigation rather than a retrospective analysis. From 1 June to 31 August 2024, all patients with suspected traumatic brain injury or suspected stroke who underwent a head CT at the largest emergency department in Norway were included. Images from six different CT scanners were pushed from the hospital’s picture archiving and communication system, known as PACS, into NeoMedSys as de-identified DICOM files, tagged by frontline users, and processed by the VIOLA-AI model within minutes.

VIOLA-AI, short for Voxels Intersecting along Orthogonal-Levels of Attention, is an ensemble of two types of three-dimensional neural networks: nnU-Net and a ViolaU-Net architecture. The model produces a binary mask demarcating any detected intracranial hemorrhage along with its volume, or a green circle when no bleed is found, and returns the result to PACS as an automated report. The mean time for VIOLA-AI to generate a result within NeoMedSys was three minutes, while the mean time from CT examination to the signed radiology report was 35 minutes, with a range stretching from three minutes to five hours.

The platform itself is built as a cloud-native hierarchical microservice with three main components. A machine learning operations layer manages the full lifecycle of the AI prototype, covering data curation, model development and fine-tuning, deployment and integration with PACS, and continuous monitoring. A database management framework handles structured and unstructured medical data with enterprise-level integrity, redundancy, failover, and scalability. A web-based frontend provides radiologists with an image viewer, annotation tools, and segmentation touch-up capabilities refined through clinical user feedback. All images are de-identified at the source, and the system adopts a zero-trust security model with server-side encryption.

During the deployment, radiologists monitored VIOLA-AI’s predictions in near real time, annotating false positives, false negatives, and segmentation inaccuracies directly on the platform. A six-year experienced radiologist performed the new annotations under the supervision of a senior neuroradiologist with more than ten years of experience, both from Oslo University Hospital. These refined annotations were incorporated into subsequent training iterations, and the team also employed ensemble methods, test-time augmentation, and threshold calibration to balance sensitivity and specificity. The model’s predictions default to a probability threshold of 0.5, which radiologists can manually adjust for clinical customization.

The training data combined public and local sources. The public INSTANCE dataset contributed 100 bleed-positive cases for initial training, while the public BHSD dataset added 152 bleed-positive training cases and 35 hold-out test cases. An initial in-house dataset of 241 bleed-positive cases, collected from local hospitals but not through NeoMedSys, supplied 206 training cases and 35 hold-out cases. A third version of the model was further refined with 40 newly annotated bleed-positive cases created by radiologists on the platform during online evaluation. The hold-out test set of 220 cases, comprising 70 bleed-positive and 150 bleed-negative cases, was strictly reserved for final evaluation and never used in training.

During the three-month study window, 2,167 patients underwent acute head CT scans across the two locations to rule out intracranial hemorrhage. Patients ranged from 6 to 96 years of age, and 36.5% were female. Radiology reports confirmed 154 cases of intracranial hemorrhage, while other pathologies encountered included non-hemorrhagic stroke, cerebral tumors, and infections. The online evaluation set comprised 304 cases, split evenly between 154 bleed-positive and 150 bleed-negative scans, with the negative cases deliberately selected as hard examples that had previously triggered false positives in earlier model versions.

The performance gains were most pronounced in the third model version, where adding the 40 new annotation cases during fine-tuning raised the AUC from 0.913 to 0.949 and sensitivity from 0.831 to 0.903 on the hold-out test. The authors note that results tested on the local hold-out dataset improved considerably, while results on the more clinically equivalent online dataset were more realistic, echoing prior reports that accuracy figures tend to drop when models enter live clinical workflows. For context, the final VIOLA-AI sensitivity of 90.3% and specificity of 89.3% sit slightly below the pooled sensitivity and specificity of 92% and 94% reported in a recent meta-analysis of AI models for detecting and annotating intracranial hemorrhage on head CT, yet they approach the reported ranges of established commercial systems such as Qure.ai’s qER, with sensitivities of 81% to 93.6%, and Aidoc, with sensitivities of 88.7% to 96.2%.

The study also illustrates where the model still struggles. False positives frequently arose from bleed-like features such as partly calcified meningiomas, motion artifacts, choroid plexus calcifications, basal ganglia calcifications, and falx or dura calcifications. False negatives tended to involve subtle or small bleeds, lesions in anatomically challenging regions, or hemorrhage types the model had difficulty recognizing; in these cases the model often identified a potential hemorrhage region but assigned a probability below the 0.5 detection threshold. The authors caution that the existence of false negatives means AI models cannot reliably produce final reports on their own, and they position the platform’s output as an assisting tool for research and quality management rather than a final decision-maker.

The team acknowledges several limitations. Communication gaps meant many head CT scans were initially not pushed to NeoMedSys, though they were pushed at a later date so all images were analyzed by VIOLA-AI. The three-month duration was chosen to capture many cases without overwhelming the volunteer staff, and the single-country setting may increase the risk of overfitting to local data characteristics. Inter-rater reliability of the annotations was not formally assessed, and a granular analysis across specific hemorrhage subtypes such as subarachnoid, subdural, and epidural bleeds remains for future work as the dataset expands. The authors plan to extend NeoMedSys to support multi-site deployments and external validation, and they argue that larger institutions with the necessary infrastructure and clinical and technological competence should invest in multidisciplinary development to bridge the gap between research and clinical practice.

Subject of Research: Real-world clinical deployment and iterative retraining of an AI model for detecting intracranial hemorrhage on head CT scans

Article Title: Examining deployment and refinement of the VIOLA-AI intracranial hemorrhage model using an interactive NeoMedSys platform

Article References: Liu, Q., Nesvold, J. E., Raaum, H., Murugesu, E., Røvang, M., Maclntosh, B. J., Bjørnerud, A., & Skogen, K. (2026). Examining deployment and refinement of the VIOLA-AI intracranial hemorrhage model using an interactive NeoMedSys platform. BMC Methods, 3(1), Article 30. https://doi.org/10.1186/s44330-026-00083-6

Image Credits: AI Generated

DOI: 10.1186/s44330-026-00083-6

Keywords: artificial intelligence, intracranial hemorrhage, head CT, radiology, NeoMedSys, VIOLA-AI, machine learning, clinical deployment, model retraining, emergency department, Oslo University Hospital, BMC Methods