Trust and stakeholder perspectives on the implementation of AI tools in clinical radiology.

OBJECTIVES: To define requirements that condition trust in artificial intelligence (AI) as clinical decision support in radiology from the perspective of various stakeholders and to explore ways to fulfil these requirements. METHODS: Semi-structured interviews were conducted with twenty-five respondents-nineteen directly involved in the development, implementation, or use of AI applications in radiology and six working with AI in other areas of healthcare. We designed the questions to explore three themes: development and use of AI, professional decision-making, and management and organizational procedures connected to AI. The transcribed interviews were analysed in an iterative coding process from open coding to theoretically informed thematic coding. RESULTS: We identified four aspects of trust that relate to reliability, transparency, quality verification, and inter-organizational compatibility. These aspects fall under the categories of substantial and procedural requirements. CONCLUSIONS: Development of appropriate levels of trust in AI in healthcare is complex and encompasses multiple dimensions of requirements. Various stakeholders will have to be involved in developing AI solutions for healthcare and radiology to fulfil these requirements. CLINICAL RELEVANCE STATEMENT: For AI to achieve advances in radiology, it must be given the opportunity to support, rather than replace, human expertise. Support requires trust. Identification of aspects and conditions for trust allows developing AI implementation strategies that facilitate advancing the field. KEY POINTS: • Dimensions of procedural and substantial demands that need to be fulfilled to foster appropriate levels of trust in AI in healthcare are conditioned on aspects related to reliability, transparency, quality verification, and inter-organizational compatibility. •Creating the conditions for trust to emerge requires the involvement of various stakeholders, who will have to compensate the problem's inherent complexity by finding and promoting well-defined solutions.

An exploratory study of the impact of CT slice thickness and inter-rater variability on anatomical accuracy of malunited distal radius models and surgical guides for corrective osteotomy.

High-resolution CT images are essential in clinical practice to accurately replicate patient anatomy for 3D virtual surgical planning and designing patient-specific surgical guides. These technologies are commonly used in corrective osteotomy of the distal radius. This study evaluated how the virtual radius models and the surgical guides' surface that is in contact with the bone vary between experienced raters. Further, the discrepancies from the reference radius of surgical guides and radius models created from CT images with slice thicknesses larger than the reference standard of 0.625mm were assessed. Maximum overlap with radius model was measured for guides, and absolute average distance error was measured for radius models. The agreement between the lower-resolution guides surface and the raters' guide surface was evaluated. The average inter-rater guide surface overlap was -0.11mm [95% CI: -0.13-0.09]. The surface of surgical guides designed on CT images with a 1mm slice thickness deviated from the reference radius within the inter-rater range (0.03mm). For slice thicknesses of 1.25mm and 1.5mm, the average guide surface overlap was 0.12mm and 0.15mm, respectively. The average inter-rater radius surface variability was 0.03mm [95% CI: 0.025-0.035]. The discrepancy from the reference of all radius models created from CT images with a slice thickness larger than the reference slice thickness was notably larger than the inter-rater variability but, excluding one case, did not exceed 0.2mm. The results suggest that 1mm CT images are suitable for surgical guide design. While 1.25mm slices are commonly used for virtual planning in hand and forearm surgery, slices larger than 1mm may approach the limit of clinical acceptability. Discrepancies in radius models were below 1mm, likely below clinical relevance.

Improving standard volar plate fixation in 3D-guided corrective osteotomy of the distal radius: evaluation of a shim instrument.

Standard volar plates often do not fit the surface of the malunited distal radius after osteotomy, necessitating an offset angle for accurate volar tilt correction. The correction can be achieved if the plate is held at the correct angle when the distal screws are locked. With the advantage of 3D surgical planning and patient-specific instruments, we developed a shim instrument to assist the surgeon in securing the plate at the intended angle when locking the distal screws, and evaluated radiological results. Five female patients aged 63-74 with dorsally angulated extra-articular malunions underwent surgery using 3D-printed guides and the shim instrument. The plate position, drilling guide alignment, screw placements, and distal radius correction on postoperative CTs were compared with the surgical plans. Errors were measured using an anatomical coordinate system, and standard 2D radiographic measures were extracted. Preoperative dorsal tilt ranged from 16° to 35°, and postoperative volar tilt from 1° to 11°. 3D analysis revealed mean absolute correction errors of 6.1° in volar tilt, 1.6° in radial inclination, and 0.6 mm in ulnar variance. The volar tilt error due to the shim instrument, indicated by the mean angle error of the distal screws to the plate, was 2.1° but varied across the five patients. Settling of the distal radius, due to tension during and after reduction, further contributed to a mean loss of 3.5° in volar tilt. The shim instrument helped with securing plates at the intended angle; however, further correction improvements should consider the tension between the fragments of osteoporotic bone.

Prediction of neurologic outcome after out-of-hospital cardiac arrest: An interpretable approach with machine learning.

Out-of-hospital cardiac arrest (OHCA) is a critical condition with low survival rates. In patients with a return of spontaneous circulation, brain injury is a leading cause of death. In this study, we propose an interpretable machine learning approach for predicting neurologic outcome after OHCA, using information available at the time of hospital admission. METHODS: The study population were 55 615 OHCA cases registered in the Swedish Cardiopulmonary Resuscitation Registry between 2010 and 2020. The dataset was split to training and validation sets (for model development) and test set (for evaluation of the final model). We used an XGBoost algorithm with stratified, repeated 10-fold cross-validation along with Optuna framework for hyperparameters tuning. The final model was trained on 10 features selected based on the importance scores and evaluated on the test set in terms of discrimination, calibration and bias-variance tradeoff. We used SHapley Additive exPlanations to address the 'black-box' model and align with eXplainable artificial intelligence. RESULTS: The final model achieved: area under the receiver operating characteristic value 0.964 (95% confidence interval (CI) [0.960-0.968]), sensitivity 0.606 (95% CI [0.573-0.634]), specificity 0.975 (95% CI [0.972-0.978]), positive predictive value (PPV) 0.664 (95% CI [0.625-0.696]), negative predictive value (NPV) 0.969 (95% CI [0.966-0.972]), macro F1 0.803 (95% CI [0.788-0.816]), and showed a very good calibration. SHAP features with the highest impact on the model's output were:'ROSC on arrival to hospital', 'Initial rhythm asystole' and 'Conscious on arrival to hospital'. CONCLUSIONS: The XGBoost machine learning model with 10 features available at the time of hospital admission showed good performance for predicting neurologic outcome after OHCA, with no apparent signs of overfitting.

Deep learning for automatic brain tumour segmentation on MRI: evaluation of recommended reporting criteria via a reproduction and replication study.

OBJECTIVES: To determine the reproducibility and replicability of studies that develop and validate segmentation methods for brain tumours on MRI and that follow established reproducibility criteria; and to evaluate whether the reporting guidelines are sufficient. METHODS: Two eligible validation studies of distinct deep learning (DL) methods were identified. We implemented the methods using published information and retraced the reported validation steps. We evaluated to what extent the description of the methods enabled reproduction of the results. We further attempted to replicate reported findings on a clinical set of images acquired at our institute consisting of high-grade and low-grade glioma (HGG, LGG), and meningioma (MNG) cases. RESULTS: We successfully reproduced one of the two tumour segmentation methods. Insufficient description of the preprocessing pipeline and our inability to replicate the pipeline resulted in failure to reproduce the second method. The replication of the first method showed promising results in terms of Dice similarity coefficient (DSC) and sensitivity (Sen) on HGG cases (DSC=0.77, Sen=0.88) and LGG cases (DSC=0.73, Sen=0.83), however, poorer performance was observed for MNG cases (DSC=0.61, Sen=0.71). Preprocessing errors were identified that contributed to low quantitative scores in some cases. CONCLUSIONS: Established reproducibility criteria do not sufficiently emphasise description of the preprocessing pipeline. Discrepancies in preprocessing as a result of insufficient reporting are likely to influence segmentation outcomes and hinder clinical utilisation. A detailed description of the whole processing chain, including preprocessing, is thus necessary to obtain stronger evidence of the generalisability of DL-based brain tumour segmentation methods and to facilitate translation of the methods into clinical practice.

Automatic brain lesion segmentation on standard magnetic resonance images: a scoping review.

OBJECTIVES: Medical image analysis practices face challenges that can potentially be addressed with algorithm-based segmentation tools. In this study, we map the field of automatic MR brain lesion segmentation to understand the clinical applicability of prevalent methods and study designs, as well as challenges and limitations in the field. DESIGN: Scoping review. SETTING: Three databases (PubMed, IEEE Xplore and Scopus) were searched with tailored queries. Studies were included based on predefined criteria. Emerging themes during consecutive title, abstract, methods and whole-text screening were identified. The full-text analysis focused on materials, preprocessing, performance evaluation and comparison. RESULTS: Out of 2990 unique articles identified through the search, 441 articles met the eligibility criteria, with an estimated growth rate of 10% per year. We present a general overview and trends in the field with regard to publication sources, segmentation principles used and types of lesions. Algorithms are predominantly evaluated by measuring the agreement of segmentation results with a trusted reference. Few articles describe measures of clinical validity. CONCLUSIONS: The observed reporting practices leave room for improvement with a view to studying replication, method comparison and clinical applicability. To promote this improvement, we propose a list of recommendations for future studies in the field.

Automatic brain lesion segmentation on standard MRIs of the human head: a scoping review protocol.

INTRODUCTION: Automatic brain lesion segmentation from medical images has great potential to support clinical decision making. Although numerous methods have been proposed, significant challenges must be addressed before they will become established in clinical and research practice. We aim to elucidate the state of the art, to provide a synopsis of competing approaches and identify contrasts between them. METHODS AND ANALYSIS: We present the background and study design of a scoping review for automatic brain lesion segmentation methods for conventional MRI according to the framework proposed by Arksey and O'Malley. We aim to identify common image processing steps as well as mathematical and computational theories implemented in these methods. We will aggregate the evidence on the efficacy and identify limitations of the approaches. Methods to be investigated work with standard MRI sequences from human patients examined for brain lesions, and are validated with quantitative measures against a trusted reference. PubMed, IEEE Xplore and Scopus will be searched using search phrases that will ensure an inclusive and unbiased overview. For matching records, titles and abstracts will be screened to ensure eligibility. Studies will be excluded if a full paper is not available or is not written in English, if non-standard MR sequences are used, if there is no quantitative validation, or if the method is not automatic. In the data charting phase, we will extract information about authors, publication details and study cohort. We expect to find information about preprocessing, segmentation and validation procedures. We will develop an analytical framework to collate, summarise and synthesise the data. ETHICS AND DISSEMINATION: Ethical approval for this study is not required since the information will be extracted from published studies. We will submit the review report to a peer-reviewed scientific journal and explore other venues for presenting the work.