MONAI Label: A framework for AI-assisted interactive labeling of 3D medical images.

The lack of annotated datasets is a major bottleneck for training new task-specific supervised machine learning models, considering that manual annotation is extremely expensive and time-consuming. To address this problem, we present MONAI Label, a free and open-source framework that facilitates the development of applications based on artificial intelligence (AI) models that aim at reducing the time required to annotate radiology datasets. Through MONAI Label, researchers can develop AI annotation applications focusing on their domain of expertise. It allows researchers to readily deploy their apps as services, which can be made available to clinicians via their preferred user interface. Currently, MONAI Label readily supports locally installed (3D Slicer) and web-based (OHIF) frontends and offers two active learning strategies to facilitate and speed up the training of segmentation algorithms. MONAI Label allows researchers to make incremental improvements to their AI-based annotation application by making them available to other researchers and clinicians alike. Additionally, MONAI Label provides sample AI-based interactive and non-interactive labeling applications, that can be used directly off the shelf, as plug-and-play to any given dataset. Significant reduced annotation times using the interactive model can be observed on two public datasets.

Deep learning for automatic segmentation of vestibular schwannoma: a retrospective study from multi-center routine MRI.

Automatic segmentation of vestibular schwannoma (VS) from routine clinical MRI has potential to improve clinical workflow, facilitate treatment decisions, and assist patient management. Previous work demonstrated reliable automatic segmentation performance on datasets of standardized MRI images acquired for stereotactic surgery planning. However, diagnostic clinical datasets are generally more diverse and pose a larger challenge to automatic segmentation algorithms, especially when post-operative images are included. In this work, we show for the first time that automatic segmentation of VS on routine MRI datasets is also possible with high accuracy. We acquired and publicly release a curated multi-center routine clinical (MC-RC) dataset of 160 patients with a single sporadic VS. For each patient up to three longitudinal MRI exams with contrast-enhanced T1-weighted (ceT1w) (n = 124) and T2-weighted (T2w) (n = 363) images were included and the VS manually annotated. Segmentations were produced and verified in an iterative process: (1) initial segmentations by a specialized company; (2) review by one of three trained radiologists; and (3) validation by an expert team. Inter- and intra-observer reliability experiments were performed on a subset of the dataset. A state-of-the-art deep learning framework was used to train segmentation models for VS. Model performance was evaluated on a MC-RC hold-out testing set, another public VS datasets, and a partially public dataset. The generalizability and robustness of the VS deep learning segmentation models increased significantly when trained on the MC-RC dataset. Dice similarity coefficients (DSC) achieved by our model are comparable to those achieved by trained radiologists in the inter-observer experiment. On the MC-RC testing set, median DSCs were 86.2(9.5) for ceT1w, 89.4(7.0) for T2w, and 86.4(8.6) for combined ceT1w+T2w input images. On another public dataset acquired for Gamma Knife stereotactic radiosurgery our model achieved median DSCs of 95.3(2.9), 92.8(3.8), and 95.5(3.3), respectively. In contrast, models trained on the Gamma Knife dataset did not generalize well as illustrated by significant underperformance on the MC-RC routine MRI dataset, highlighting the importance of data variability in the development of robust VS segmentation models. The MC-RC dataset and all trained deep learning models were made available online.

Symptoms before and after COVID-19: a population and case-control study using prospective data.

BACKGROUND: Some individuals experience prolonged illness after acute COVID-19. We assessed whether pre-infection symptoms affected post-COVID illness duration. METHODS: Survival analysis was performed in adults (n=23 452) with community-managed SARC-CoV-2 infection prospectively self-logging data through the ZOE COVID Symptom Study app, at least weekly, from 8 weeks before to 12 weeks after COVID-19 onset, conditioned on presence versus absence of baseline symptoms (4-8 weeks before COVID-19). A case-control study was performed in 1350 individuals with long illness (≥8 weeks, 906 [67.1%] with illness ≥12 weeks), matched 1:1 (for age, sex, body mass index, testing week, prior infection, vaccination, smoking, index of multiple deprivation) with 1350 individuals with short illness (<4 weeks). Baseline symptoms were compared between the two groups; and against post-COVID symptoms. RESULTS: Individuals reporting baseline symptoms had longer post-COVID symptom duration (from 10 to 15 days) with baseline fatigue nearly doubling duration. Two-thirds (910 of 1350 [67.4%]) of individuals with long illness were asymptomatic beforehand. However, 440 (32.6%) had baseline symptoms, versus 255 (18.9%) of 1350 individuals with short illness (p<0.0001). Baseline symptoms increased the odds ratio for long illness (2.14 [CI: 1.78; 2.57]). Prior comorbidities were more common in individuals with long versus short illness. In individuals with long illness, baseline symptomatic (versus asymptomatic) individuals were more likely to be female, younger, and have prior comorbidities; and baseline and post-acute symptoms and symptom burden correlated strongly. CONCLUSIONS: Individuals experiencing symptoms before COVID-19 have longer illness duration and increased odds of long illness. However, many individuals with long illness are well before SARS-CoV-2 infection.

Prospective development study of the Versius Surgical System for use in transoral robotic surgery: an IDEAL stage 1/2a first in human and initial case series experience.

PURPOSE: Transoral robotic surgery is well established in the treatment paradigm of oropharyngeal pathology. The Versius Surgical System (CMR Surgical) is a robotic platform in clinical use in multiple specialities but is currently untested in the head and neck. This study utilises the IDEAL framework of surgical innovation to prospectively evaluate and report a first in human clinical experience and single centre case series of transoral robotic surgery (TORS) with Versius. METHODS: Following IDEAL framework stages 1 and 2a, the study evaluated Versius to perform first in human TORS before transitioning from benign to malignant cases. Iterative adjustments were made to system setup, instrumentation, and technique, recorded in accordance with IDEAL recommendations. Evaluation criteria included successful procedure completion, setup time, operative time, complications, and subjective impressions. Further evaluation of the system to perform four-arm surgery was conducted. RESULTS: 30 TORS procedures were successfully completed (15 benign, 15 malignant) without intraoperative complication or conversion to open surgery. Setup time significantly decreased over the study period. Instrumentation challenges were identified, urging the need for TORS-specific instruments. The study introduced four-arm surgery, showcasing Versius' unique capabilities, although limitations in distal access were observed. CONCLUSIONS: TORS is feasible with the Versius Surgical System. The development of TORS-specific instruments would benefit performance and wider adoption of the system. 4-arm surgery is possible however further evaluation is required. Multicentre evaluation (IDEAL stage 2b) is recommended.

Optimising brain age estimation through transfer learning: A suite of pre-trained foundation models for improved performance and generalisability in a clinical setting.

Estimated age from brain MRI data has emerged as a promising biomarker of neurological health. However, the absence of large, diverse, and clinically representative training datasets, along with the complexity of managing heterogeneous MRI data, presents significant barriers to the development of accurate and generalisable models appropriate for clinical use. Here, we present a deep learning framework trained on routine clinical data (N up to 18,890, age range 18-96 years). We trained five separate models for accurate brain age prediction (all with mean absolute error ≤4.0 years, R2 ≥ .86) across five different MRI sequences (T2 -weighted, T2 -FLAIR, T1 -weighted, diffusion-weighted, and gradient-recalled echo T2 *-weighted). Our trained models offer dual functionality. First, they have the potential to be directly employed on clinical data. Second, they can be used as foundation models for further refinement to accommodate a range of other MRI sequences (and therefore a range of clinical scenarios which employ such sequences). This adaptation process, enabled by transfer learning, proved effective in our study across a range of MRI sequences and scan orientations, including those which differed considerably from the original training datasets. Crucially, our findings suggest that this approach remains viable even with limited data availability (as low as N = 25 for fine-tuning), thus broadening the application of brain age estimation to more diverse clinical contexts and patient populations. By making these models publicly available, we aim to provide the scientific community with a versatile toolkit, promoting further research in brain age prediction and related areas.

New robotic platform for transoral robotic surgery: an IDEAL stage 0 study.

Objectives: This study aims to assess the feasibility to perform transoral robotic surgery (TORS) with a new robotic platform, the Versius Surgical System (CMR Surgical, UK) in a preclinical cadaveric setting in accordance to stage 0 of the IDEAL-D framework. Design: IDEAL stage 0 preclinical assessment of the Versius Robotic System in TORS in human cadavers. Setting: All procedures were performed in a simulated operating theatre environment at a UK surgical training centre. Participants: 11 consultant head and neck surgeons from the UK, mainland Europe and the USA took part in TORS procedures on six human cadavers. Interventions: 3 key index procedures were assessed that represent the core surgical workload of TORS: lateral oropharyngectomy, tongue base resection and partial supraglottic laryngectomy. Main outcome measures: The primary outcome was the successful completion of each surgical procedure. Secondary outcomes included the optimisation of system setup, instrumentation and surgeon-reported outcomes for feasibility of each component procedural step. Results: 33 cadaveric procedures were performed and 32 were successfully completed. One supraglottic laryngectomy was not fully completed due to issues dividing the epiglottic cartilage with available instrumentation. Surgeon-reported outcomes met the minimal level of feasibility in all procedures and a consensus that it is feasible to perform TORS with Versius was reached. Available instrumentation was not representative of other robotic platforms used in TORS and further instrument optimisation is recommended before wider dissemination. Conclusions: It is feasible to perform TORS with the Versius Surgical System (CMR Surgical) within a pre-clinical cadaveric setting. Clinical evaluation is needed and appropriate with the system. Further instrument development and optimisation is desirable.

Unpaired intra-operative OCT (iOCT) video super-resolution with contrastive learning.

Regenerative therapies show promise in reversing sight loss caused by degenerative eye diseases. Their precise subretinal delivery can be facilitated by robotic systems alongside with Intra-operative Optical Coherence Tomography (iOCT). However, iOCT's real-time retinal layer information is compromised by inferior image quality. To address this limitation, we introduce an unpaired video super-resolution methodology for iOCT quality enhancement. A recurrent network is proposed to leverage temporal information from iOCT sequences, and spatial information from pre-operatively acquired OCT images. Additionally, a patchwise contrastive loss enables unpaired super-resolution. Extensive quantitative analysis demonstrates that our approach outperforms existing state-of-the-art iOCT super-resolution models. Furthermore, ablation studies showcase the importance of temporal aggregation and contrastive loss in elevating iOCT quality. A qualitative study involving expert clinicians also confirms this improvement. The comprehensive evaluation demonstrates our method's potential to enhance the iOCT image quality, thereby facilitating successful guidance for regenerative therapies.

Assessing the optimal MRI descriptors to diagnose Ménière's disease and the added value of analysing the vestibular aqueduct.

OBJECTIVES: To evaluate the diagnostic performance and reliability of MRI descriptors used for the detection of Ménière's disease (MD) on delayed post-gadolinium MRI. To determine which combination of descriptors should be optimally applied and whether analysis of the vestibular aqueduct (VA) contributes to the diagnosis. MATERIALS AND METHODS: This retrospective single centre case-control study evaluated delayed post-gadolinium MRI of patients with Ménièriform symptoms examined consecutively between Dec 2017 and March 2023. Two observers evaluated 17 MRI descriptors of MD and quantified perilymphatic enhancement (PLE) in the cochlea. Definite MD ears according to the 2015 Barany Society criteria were compared to control ears. Cohen's kappa and diagnostic odds ratio (DORs) were calculated for each descriptor. Forward stepwise logistic regression determined which combination of MRI descriptors would best predict MD ears, and the area under the receiver operating characteristic curve for this model was measured. RESULTS: A total of 227 patients (mean age 48.3 ± 14.6, 99 men) with 96 definite MD and 78 control ears were evaluated. The presence of saccular abnormality (absent, as large as or confluent with the utricle) performed best with a DOR of 292.6 (95% confidence interval (CI), 38.305-2235.058). All VA descriptors demonstrated excellent reliability and with DORs of 7.761 (95% CI, 3.517-17.125) to 18.1 (95% CI, 8.445-39.170). Combining these saccular abnormalities with asymmetric cochlear PLE and an incompletely visualised VA correctly classified 90.2% of cases (sensitivity 84.4%, specificity 97.4%, AUC 0.938). CONCLUSION: Either absent, enlarged or confluent saccules are the best predictors of MD. Incomplete visualisation of the VA adds value to the diagnosis. CLINICAL RELEVANCE STATEMENT: A number of different MRI descriptors have been proposed for the diagnosis of Ménière's disease, but by establishing the optimally performing MRI features and highlighting new useful descriptors, there is an opportunity to improve the diagnostic performance of Ménière's disease imaging. KEY POINTS: • A comprehensive range of existing and novel vestibular aqueduct delayed post-gadolinium MRI descriptors were compared for their diagnostic performance in Ménière's disease. • Saccular abnormality (absent, confluent with or larger than the utricle) is a reliable descriptor and is the optimal individual MRI predictor of Ménière's disease. • The presence of this saccule descriptor or asymmetric perilymphatic enhancement and incomplete vestibular aqueduct visualisation will optimise the MRI diagnosis of Ménière's disease.

A Dempster-Shafer Approach to Trustworthy AI With Application to Fetal Brain MRI Segmentation.

Deep learning models for medical image segmentation can fail unexpectedly and spectacularly for pathological cases and images acquired at different centers than training images, with labeling errors that violate expert knowledge. Such errors undermine the trustworthiness of deep learning models for medical image segmentation. Mechanisms for detecting and correcting such failures are essential for safely translating this technology into clinics and are likely to be a requirement of future regulations on artificial intelligence (AI). In this work, we propose a trustworthy AI theoretical framework and a practical system that can augment any backbone AI system using a fallback method and a fail-safe mechanism based on Dempster-Shafer theory. Our approach relies on an actionable definition of trustworthy AI. Our method automatically discards the voxel-level labeling predicted by the backbone AI that violate expert knowledge and relies on a fallback for those voxels. We demonstrate the effectiveness of the proposed trustworthy AI approach on the largest reported annotated dataset of fetal MRI consisting of 540 manually annotated fetal brain 3D T2w MRIs from 13 centers. Our trustworthy AI method improves the robustness of four backbone AI models for fetal brain MRIs acquired across various centers and for fetuses with various brain abnormalities.

The impact of the size and angle of the cochlear basal turn on translocation of a pre-curved mid-scala cochlear implant electrode.

Scalar translocation is a severe form of intra-cochlear trauma during cochlear implant (CI) electrode insertion. This study explored the hypothesis that the dimensions of the cochlear basal turn and orientation of its inferior segment relative to surgically relevant anatomical structures influence the scalar translocation rates of a pre-curved CI electrode. In a cohort of 40 patients implanted with the Advanced Bionics Mid-Scala electrode array, the scalar translocation group (40%) had a significantly smaller mean distance A of the cochlear basal turn (p < 0.001) and wider horizontal angle between the inferior segment of the cochlear basal turn and the mastoid facial nerve (p = 0.040). A logistic regression model incorporating distance A (p = 0.003) and horizontal facial nerve angle (p = 0.017) explained 44.0-59.9% of the variance in scalar translocation and correctly classified 82.5% of cases. Every 1mm decrease in distance A was associated with a 99.2% increase in odds of translocation [95% confidence interval 80.3%, 100%], whilst every 1-degree increase in the horizontal facial nerve angle was associated with an 18.1% increase in odds of translocation [95% CI 3.0%, 35.5%]. The study findings provide an evidence-based argument for the development of a navigation system for optimal angulation of electrode insertion during CI surgery to reduce intra-cochlear trauma.

Glioblastoma and Radiotherapy: a multi-center AI study for Survival Predictions from MRI (GRASP study).

BACKGROUND: The aim was to predict survival of glioblastoma at eight months after radiotherapy (a period allowing for completing a typical course of adjuvant temozolomide), by applying deep learning to the first brain MRI after radiotherapy completion. METHODS: Retrospective and prospective data were collected from 206 consecutive glioblastoma, IDH-wildtype patients diagnosed between March 2014-February 2022 across 11 UK centers. Models were trained on 158 retrospective patients from three centers. Holdout test sets were retrospective (n=19; internal validation), and prospective (n=29; external validation from eight distinct centers).Neural network branches for T2-weighted and contrast-enhanced T1-weighted inputs were concatenated to predict survival. A non-imaging branch (demographics/MGMT/treatment data) was also combined with the imaging model. We investigated the influence of individual MR sequences; non-imaging features; and weighted dense blocks pretrained for abnormality detection. RESULTS: The imaging model outperformed the non-imaging model in all test sets (area under the receiver-operating characteristic curve, AUC p=0.038) and performed similarly to a combined imaging/non-imaging model (p>0.05). Imaging, non-imaging, and combined models applied to amalgamated test sets gave AUCs of 0.93, 0.79, and 0.91. Initializing the imaging model with pretrained weights from 10,000s of brain MRIs improved performance considerably (amalgamated test sets without pretraining 0.64; p=0.003). CONCLUSIONS: A deep learning model using MRI images after radiotherapy, reliably and accurately determined survival of glioblastoma. The model serves as a prognostic biomarker identifying patients who will not survive beyond a typical course of adjuvant temozolomide, thereby stratifying patients into those who might require early second-line or clinical trial treatment.

Multi-source multi-modal markers for Bayesian Networks: Application to the extremely preterm born brain.

The preterm phenotype results from the interplay of multiple disorders affecting the brain and cognitive outcomes. Accurately characterising these interactions can reveal prematurity markers. Bayesian Networks (BNs) are powerful tools to disentangle these relationships, as they inherently measure associations between variables while mitigating confounding factors. We present Modified PC-HC (MPC-HC), a Bayesian Network (BN) structural learning algorithm. MPC-HC employs statistical testing and search-and-score techniques to explore equivalent classes. We employ MPC-HC to estimate BNs for extremely preterm (EP) young adults and full-term controls. Using MRI measurements and cognitive performance markers, we investigate predictive relationships and mutual influences through predictions and sensitivity analysis. We assess the confidence in the estimated BN structures using bootstrapping. Furthermore, MPC-HC's validation involves assessing its ability to recover benchmark BN structures. MPC-HC achieves an average prediction accuracy of 72.5% compared to 62.5% of PC, 64.5% of MMHC, and 71.5% of HC, while it outperforms PC, MMHC, and HC algorithms in reconstructing the true structure of benchmark BNs. The sensitivity analysis shows that MRI measurements mainly affect EP cognitive scores. Our work has two key contributions: first, the introduction and validation of a new BN structure learning method. Second, demonstrating the potential of BNs in modelling variable relationships, predicting variables of interest, modelling uncertainty, and evaluating how variables impact each other. Finally, we demonstrate this by characterising complex phenotypes, such as preterm birth, and discovering results consistent with literature findings.

Acquisition-invariant brain MRI segmentation with informative uncertainties.

Combining multi-site data can strengthen and uncover trends, but is a task that is marred by the influence of site-specific covariates that can bias the data and, therefore, any downstream analyses. Post-hoc multi-site correction methods exist but have strong assumptions that often do not hold in real-world scenarios. Algorithms should be designed in a way that can account for site-specific effects, such as those that arise from sequence parameter choices, and in instances where generalisation fails, should be able to identify such a failure by means of explicit uncertainty modelling. This body of work showcases such an algorithm that can become robust to the physics of acquisition in the context of segmentation tasks while simultaneously modelling uncertainty. We demonstrate that our method not only generalises to complete holdout datasets, preserving segmentation quality but does so while also accounting for site-specific sequence choices, which also allows it to perform as a harmonisation tool.

Diversity of symptom phenotypes in SARS-CoV-2 community infections observed in multiple large datasets.

Variability in case severity and in the range of symptoms experienced has been apparent from the earliest months of the COVID-19 pandemic. From a clinical perspective, symptom variability might indicate various routes/mechanisms by which infection leads to disease, with different routes requiring potentially different treatment approaches. For public health and control of transmission, symptoms in community cases were the prompt upon which action such as PCR testing and isolation was taken. However, interpreting symptoms presents challenges, for instance, in balancing the sensitivity and specificity of individual symptoms with the need to maximise case finding, whilst managing demand for limited resources such as testing. For both clinical and transmission control reasons, we require an approach that allows for the possibility of distinct symptom phenotypes, rather than assuming variability along a single dimension. Here we address this problem by bringing together four large and diverse datasets deriving from routine testing, a population-representative household survey and participatory smartphone surveillance in the United Kingdom. Through the use of cutting-edge unsupervised classification techniques from statistics and machine learning, we characterise symptom phenotypes among symptomatic SARS-CoV-2 PCR-positive community cases. We first analyse each dataset in isolation and across age bands, before using methods that allow us to compare multiple datasets. While we observe separation due to the total number of symptoms experienced by cases, we also see a separation of symptoms into gastrointestinal, respiratory and other types, and different symptom co-occurrence patterns at the extremes of age. In this way, we are able to demonstrate the deep structure of symptoms of COVID-19 without usual biases due to study design. This is expected to have implications for the identification and management of community SARS-CoV-2 cases and could be further applied to symptom-based management of other diseases and syndromes.

3D vs. 2D simulated fetoscopy for spina bifida repair: a quantitative motion analysis.

3D imaging technology is becoming more prominent every day. However, more validation is needed to understand the actual benefit of 3D versus conventional 2D vision. This work quantitatively investigates whether experts benefit from 3D vision during minimally invasive fetoscopic spina bifida (fSB) repair. A superiority study was designed involving one expert team ([Formula: see text] procedures prior) who performed six 2D and six 3D fSB repair simulations in a high-fidelity animal training model, using 3-port access. The 6D motion of the instruments was recorded. Among the motion metrics are total path length, smoothness, maximum speed, the modified Spectral Arc Length (SPARC), and Log Dimensionless Jerk (LDLJ). The primary clinical outcome is operation time (power 90%, 5% significance) using Sealed Envelope Ltd. 2012. Secondary clinical outcomes are water tightness of the repair, CO[Formula: see text] insufflation volume, and OSATS score. Findings show that total path length and LDLJ are considerably different. Operation time during 3D vision was found to be significantly shorter compared to 2D vision ([Formula: see text] vs. [Formula: see text] min; p [Formula: see text] 0.026). These results suggest enhanced performance with 3D vision during interrupted suturing in fetoscopic SBA repair. To confirm these results, a larger-scale follow-up study involving multiple experts and novice surgeons is recommended.

Medical device regulatory challenges in the UK are affecting innovation and its potential benefits.

The increase in regulatory challenges on medical technology developed and deployed in the UK is having a negative impact on innovation. In this paper we show how the limited capacity of Approved and Notified Bodies is one more barrier in the innovation pipeline, that could push more teams to consider applying for FDA approval instead of UKCA marking, potentially limiting how much our patients benefit from the world-leading research undertaken in UK universities.

Lightfield hyperspectral imaging in neuro-oncology surgery: an IDEAL 0 and 1 study.

Introduction: Hyperspectral imaging (HSI) has shown promise in the field of intra-operative imaging and tissue differentiation as it carries the capability to provide real-time information invisible to the naked eye whilst remaining label free. Previous iterations of intra-operative HSI systems have shown limitations, either due to carrying a large footprint limiting ease of use within the confines of a neurosurgical theater environment, having a slow image acquisition time, or by compromising spatial/spectral resolution in favor of improvements to the surgical workflow. Lightfield hyperspectral imaging is a novel technique that has the potential to facilitate video rate image acquisition whilst maintaining a high spectral resolution. Our pre-clinical and first-in-human studies (IDEAL 0 and 1, respectively) demonstrate the necessary steps leading to the first in-vivo use of a real-time lightfield hyperspectral system in neuro-oncology surgery. Methods: A lightfield hyperspectral camera (Cubert Ultris ×50) was integrated in a bespoke imaging system setup so that it could be safely adopted into the open neurosurgical workflow whilst maintaining sterility. Our system allowed the surgeon to capture in-vivo hyperspectral data (155 bands, 350-1,000 nm) at 1.5 Hz. Following successful implementation in a pre-clinical setup (IDEAL 0), our system was evaluated during brain tumor surgery in a single patient to remove a posterior fossa meningioma (IDEAL 1). Feedback from the theater team was analyzed and incorporated in a follow-up design aimed at implementing an IDEAL 2a study. Results: Focusing on our IDEAL 1 study results, hyperspectral information was acquired from the cerebellum and associated meningioma with minimal disruption to the neurosurgical workflow. To the best of our knowledge, this is the first demonstration of HSI acquisition with 100+ spectral bands at a frame rate over 1Hz in surgery. Discussion: This work demonstrated that a lightfield hyperspectral imaging system not only meets the design criteria and specifications outlined in an IDEAL-0 (pre-clinical) study, but also that it can translate into clinical practice as illustrated by a successful first in human study (IDEAL 1). This opens doors for further development and optimisation, given the increasing evidence that hyperspectral imaging can provide live, wide-field, and label-free intra-operative imaging and tissue differentiation.

Latent Transformer Models for out-of-distribution detection.

Any clinically-deployed image-processing pipeline must be robust to the full range of inputs it may be presented with. One popular approach to this challenge is to develop predictive models that can provide a measure of their uncertainty. Another approach is to use generative modelling to quantify the likelihood of inputs. Inputs with a low enough likelihood are deemed to be out-of-distribution and are not presented to the downstream predictive model. In this work, we evaluate several approaches to segmentation with uncertainty for the task of segmenting bleeds in 3D CT of the head. We show that these models can fail catastrophically when operating in the far out-of-distribution domain, often providing predictions that are both highly confident and wrong. We propose to instead perform out-of-distribution detection using the Latent Transformer Model: a VQ-GAN is used to provide a highly compressed latent representation of the input volume, and a transformer is then used to estimate the likelihood of this compressed representation of the input. We demonstrate this approach can identify images that are both far- and near- out-of-distribution, as well as provide spatial maps that highlight the regions considered to be out-of-distribution. Furthermore, we find a strong relationship between an image's likelihood and the quality of a model's segmentation on it, demonstrating that this approach is viable for filtering out unsuitable images.