Expert-level Performance Research Articles

TVnet: automated global analysis of tricuspid valve plane motion in CMR long-axis cines with residual neural networks for assessment of right ventricular function

Abstract Funding Acknowledgements Type of funding sources: Foundation. Main funding source(s): Clarendon Fund, John Fell Oxford University Press Research Fund, Oxford BHF Centre of Research Excellence (RE/18/3/34214), Alison Brading Memorial Graduate Scholarship in Medical Science, National Institute for Health Research Oxford Biomedical Research Centre, National Institutes of Health (R01HL144706). Background Right ventricular (RV) function evaluation is an integral part of comprehensive cardiac assessment, including for pulmonary hypertension, congenital heart disease and arrhythmogenic RV cardiomyopathy (ARVC) [1]. It is commonly assessed by measuring tricuspid annular plane systolic excursion (TAPSE) and peak systolic velocity (RV s’) on echocardiography [2]. However, it is highly sensitive to imaging window and small changes in the beam angle, limiting reliability [3]. Cardiovascular magnetic resonance (CMR) is the imaging gold-standard for assessing RV structure and function, and is highly reproducible. CMR can assess tricuspid valve (TV) motion using four-chamber (4Ch) and RV two-chamber (2Ch) cines, with high diagnostic performance when compared against single-plane analysis [4]. However, manual placement of the TV insertion points is highly time-consuming for routine clinical workflows. TVnet, a deep-learning framework for automatically tracking the TV in 4Ch cines [5] has been recently validated, but without the orthogonal plane (RV 2Ch) which is helpful to more reliably characterise TV motion. Purpose We further extend TVnet to automatically track RV 2Ch cines and derive analysis of global TV motion parameters (global TAPSE and RV s’) on par with expert level performance. Methods 74 patients undergoing CMR (1.5T Siemens MR scanner) with 4Ch and RV 2Ch views were retrospectively included in this ethically-approved study. The patients had the following cardiovascular conditions: myocardial infarction (n=43), ARVC (n=28) and Takotsubo cardiomyopathy (n=3). The dual-stage deep-learning pipeline with a residual neural network backbone [5, 6] (Figure 1A) was trained using 69 patient datasets and 15 patients were randomly chosen for testing. The TVnet trained on 4Ch cines [5] was used to automatically annotate the 4Ch cines of the testing set for global analysis comparison. For manual reference, the software Segment [7] was used to manually annotate the TV insertion points in all imaging data (1,865 RV 2Ch images, 375 4Ch images). The global TAPSE and RV s’ were derived as the mean perpendicular motion from the end-diastolic plane from both chamber views (Figure 1B). Results TVnet achieved a fast processing accuracy (<1 second per cine) and successfully tracked the TV insertion points from the RV 2Ch cine with a plane tracking error of 0.08±0.83 mm. The integrated pipeline yielded an excellent clinical-metric agreement with the manual reference (Figure 2) for both global TAPSE (error=0.42±0.68 mm, ICC=0.95) and RV s’ (error=-0.10±0.71 cm/s, ICC=0.85). Conclusion TVnet demonstrated excellent performance in both tracking the TV insertion points in RV 2Ch cines and deriving global TAPSE and RV s’ compared to manual reference. TVnet can eventually provide a complete automatic inline analysis of TV plane motion for a fast, reliable and reproducible assessment of RV function in routine clinical workflows. (A) TVnet pipeline (B) Metric derivationCorrelation and Bland-Altman plots

Reducing systematic review burden using Deduklick: a novel, automated, reliable, and explainable deduplication algorithm to foster medical research

BackgroundIdentifying and removing reference duplicates when conducting systematic reviews (SRs) remain a major, time-consuming issue for authors who manually check for duplicates using built-in features in citation managers. To address issues related to manual deduplication, we developed an automated, efficient, and rapid artificial intelligence-based algorithm named Deduklick. Deduklick combines natural language processing algorithms with a set of rules created by expert information specialists. MethodsDeduklick’s deduplication uses a multistep algorithm of data normalization, calculates a similarity score, and identifies unique and duplicate references based on metadata fields, such as title, authors, journal, DOI, year, issue, volume, and page number range. We measured and compared Deduklick’s capacity to accurately detect duplicates with the information specialists’ standard, manual duplicate removal process using EndNote on eight existing heterogeneous datasets. Using a sensitivity analysis, we manually cross-compared the efficiency and noise of both methods.DiscussionDeduklick achieved average recall of 99.51%, average precision of 100.00%, and average F1 score of 99.75%. In contrast, the manual deduplication process achieved average recall of 88.65%, average precision of 99.95%, and average F1 score of 91.98%. Deduklick achieved equal to higher expert-level performance on duplicate removal. It also preserved high metadata quality and drastically reduced time spent on analysis. Deduklick represents an efficient, transparent, ergonomic, and time-saving solution for identifying and removing duplicates in SRs searches. Deduklick could therefore simplify SRs production and represent important advantages for scientists, including saving time, increasing accuracy, reducing costs, and contributing to quality SRs.

Self-evolving vision transformer for chest X-ray diagnosis through knowledge distillation

Although deep learning-based computer-aided diagnosis systems have recently achieved expert-level performance, developing a robust model requires large, high-quality data with annotations that are expensive to obtain. This situation poses a conundrum that annually-collected chest x-rays cannot be utilized due to the absence of labels, especially in deprived areas. In this study, we present a framework named distillation for self-supervision and self-train learning (DISTL) inspired by the learning process of the radiologists, which can improve the performance of vision transformer simultaneously with self-supervision and self-training through knowledge distillation. In external validation from three hospitals for diagnosis of tuberculosis, pneumothorax, and COVID-19, DISTL offers gradually improved performance as the amount of unlabeled data increase, even better than the fully supervised model with the same amount of labeled data. We additionally show that the model obtained with DISTL is robust to various real-world nuisances, offering better applicability in clinical setting.

Automatic segmentation of multitype retinal fluid from optical coherence tomography images using semisupervised deep learning network

Background/aimsTo develop and validate a deep learning model for automated segmentation of multitype retinal fluid using optical coherence tomography (OCT) images.MethodsWe retrospectively collected a total of 2814 completely anonymised OCT...

Convolutional neural network assistance significantly improves dermatologists’ diagnosis of cutaneous tumours using clinical images

BackgroundConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation. MethodsA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18 ≤ 10 years; 5/18＞10 years) in interpreting the 400 clinical images with or without CNN assistance. ResultsThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without. ConclusionsCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.

Assessment of deep convolutional neural network models for mandibular fracture detection in panoramic radiographs

The aim of this study was to develop automated models for the identification and detection of mandibular fractures in panoramic radiographs using convolutional neural network (CNN) algorithms. A total of 1710 panoramic radiograph images from the years 2016 to 2020, including 855 images containing mandibular fractures, were obtained retrospectively from the regional trauma centre. CNN-based classification models, DenseNet-169 and ResNet-50, were fabricated to identify fractures in the radiographic images. The CNN-based object detection models Faster R-CNN and YOLOv5 were trained to automate the placement of the bounding boxes to detect fractures in the radiographic images. The performance of the models was evaluated on a hold-out test set and also by comparison with residents in oral and maxillofacial surgery and oral and maxillofacial surgeons (experts) on a 100-image subset. The binary classification performance of the models achieved promising results with an area under the receiver operating characteristics curve (AUC), sensitivity, and specificity of 100%. The detection performance of the models achieved an AUC of approximately 90%. When compared with the accuracy of clinician observers, the identification performance of the models outperformed even an expert-level classification. In conclusion, CNN-based models identified mandibular fractures above expert-level performance. It is expected that these models will be used as an aid to improve clinician performance, with aided resident performance approximating that of expert level.

Development of a Simulation Surgical Cricothyroidotomy Curriculum for Novice Providers: A Learning Curve Study.

Airway obstruction is the third most common cause of preventable death on the battlefield, accounting for 1%-2% of total combat fatalities. No previous surgical cricothyroidotomy (SC) studies have analyzed the learning curve required to obtain proficiency despite being studied in numerous other surgical technique training experiments. The aims of this study were to establish expert SC performance criteria, develop a novel standardized SC curriculum, and determine the necessary number of practice iterations required by a novice to reach this pre-determined performance goal. A standardized checklist and SC performance standards were established based on the performance of 12 board certified Military Health System surgeons with prior experience on performing a SC using a simulated trauma mannequin. Expert-level criteria were defined as a SC time to completion of 40 s or less and checklist score of at least 9/10, including all critical steps. Study subjects included 89 novice providers (54 active-duty first- and second-year medical students and 35 Navy corpsmen). Subjects received instruction on performing a SC using the principles of mastery learning and performed a final test of SC proficiency on a trauma mannequin within a realistic simulated MEDEVAC helicopter. The total number of subject practice attempts, checklist scores, and time to completion were measured and/or blindly scored. Learning curve and exponential plateau equations were used to characterize their improvement in mean time to SC completion and checklist scores. Mean pre-test knowledge scores for the entire group were 11.8 ± 3.1 out of 24 points. Total mean practice learning plateaued at checklist scores of 9.9/10 after 7 iterations and at a mean completion time of 30.4 s after 10 iterations. During the final test performance in the helicopter, 67.4% of subjects achieved expert-level performance on the first attempt. All subjects achieved expert-level performance by the end of two additional attempts. While a significantly larger proportion of medical students (79.9%) successfully completed the helicopter test on the first attempt compared to corpsmen (54.3%), there were no statistically significant differences in mean SC completion times and checklist scores between both groups (P > 0.05). Medical students performed a SC only 1.3 s faster and scored only 0.16 points higher than corpsmen. The effect size for differences were small to negligible (Cohen's d range 0.18-0.33 for SC completion time; Cohen's d range 0.45-0.46 for checklist scores). This study successfully defined SC checklist scores and completion times based on the performance of experienced surgeons on a simulator. Using these criteria and the principles of mastery learning, novices with little knowledge and experience in SC were successfully trained to the level of experienced providers. All subjects met performance targets after training and overall performance plateaued after approximately seven iterations. Over two-thirds of subjects achieved the performance target on the first testing attempt in a simulated helicopter environment. Performance was comparable between medical student and corpsmen subgroups. Further research will assess the durability of maintaining SC skills and the timing for introducing refresher courses after initial skill acquisition.

Artificial intelligence in clinical research of cancers.

Several factors, including advances in computational algorithms, the availability of high-performance computing hardware, and the assembly of large community-based databases, have led to the extensive application of Artificial Intelligence (AI) in the biomedical domain for nearly 20 years. AI algorithms have attained expert-level performance in cancer research. However, only a few AI-based applications have been approved for use in the real world. Whether AI will eventually be capable of replacing medical experts has been a hot topic. In this article, we first summarize the cancer research status using AI in the past two decades, including the consensus on the procedure of AI based on an ideal paradigm and current efforts of the expertise and domain knowledge. Next, the available data of AI process in the biomedical domain are surveyed. Then, we review the methods and applications of AI in cancer clinical research categorized by the data types including radiographic imaging, cancer genome, medical records, drug information and biomedical literatures. At last, we discuss challenges in moving AI from theoretical research to real-world cancer research applications and the perspectives toward the future realization of AI participating cancer treatment.

EANDC: An explainable attention network based deep adaptive clustering model for mental health treatment

Internet-delivered Psychological Treatment (IDPT) has been shown to be an effective method for improving psychological disorders. Natural language processing (NLP) requires an appropriate set of linguistic features for word representation and emotion segmentation. For psychological applications, models must be trained on extensive and diverse datasets to achieve expert-level performance. Labeling psychological texts authorized by patients is challenging because emotional biases can lead to incorrect segmentation of emotions and labeling emotional data is time consuming. In this paper, we propose an assistance tool for psychologists to explore the emotional aspects of mentally ill individuals. We first use an NLP-based method to create emotional lexicon embeddings, and then apply attention-based deep clustering. The learned representation is then used to visualize the emotional aspect of the text authorized by patients. We expand the patient authored text using synonymous semantic expansion. A latent semantic representation based on context is clustered using EANDC, which is a Explainable Attention Network-based Deep adaptive Clustering model. We use similarity metrics to select a subset of the text and then improve the explainability of learning using a curriculum-based optimization method. The experimental results show that synonym expansion based on the emotion lexicon increases accuracy without affecting the results. The attention method with bidirectional LSTM architecture achieved 0.81 ROC in a blind test. The self-learning based embedding then visualizes the weighted attention words and helps the psychiatrist to improve his explanatory power of the qualitative match for clinical notes and the remedy. The method helps in labeling text and improves the recognition rate of symptoms of mental disorders.

Explainability of Convolutional Neural Networks for Dermatological Diagnosis

Background Convolutional neural networks (CNNs) are regarded as state-of-the-art artificial intelligence (AI) tools for dermatological diagnosis, and they have been shown to achieve expert-level performance when trained on a representative dataset. CNN explainability is a key factor to adopting such techniques in practice and can be achieved using attention maps of the network. However, evaluation of CNN explainability has been limited to visual assessment and remains qualitative, subjective, and time consuming. Objective This study aimed to provide a framework for an objective quantitative assessment of the explainability of CNNs for dermatological diagnosis benchmarks. Methods We sourced 566 images available under the Creative Commons license from two public datasets—DermNet NZ and SD-260, with reference diagnoses of acne, actinic keratosis, psoriasis, seborrheic dermatitis, viral warts, and vitiligo. Eight dermatologists with teledermatology expertise annotated each clinical image with a diagnosis, as well as diagnosis-supporting characteristics and their localization. A total of 16 supporting visual characteristics were selected, including basic terms such as macule, nodule, papule, patch, plaque, pustule, and scale, and additional terms such as closed comedo, cyst, dermatoglyphic disruption, leukotrichia, open comedo, scar, sun damage, telangiectasia, and thrombosed capillary. The resulting dataset consisted of 525 images with three rater annotations for each. Explainability of two fine-tuned CNN models, ResNet-50 and EfficientNet-B4, was analyzed with respect to the reference explanations provided by the dermatologists. Both models were pretrained on the ImageNet natural image recognition dataset and fine-tuned using 3214 images of the six target skin conditions obtained from an internal clinical dataset. CNN explanations were obtained as activation maps of the models through gradient-weighted class-activation maps. We computed the fuzzy sensitivity and specificity of each characteristic attention map with regard to both the fuzzy gold standard characteristic attention fusion masks and the fuzzy union of all characteristics. Results On average, explainability of EfficientNet-B4 was higher than that of ResNet-50 in terms of sensitivity for 13 of 16 supporting characteristics, with mean values of 0.24 (SD 0.07) and 0.16 (SD 0.05), respectively. However, explainability was lower in terms of specificity, with mean values of 0.82 (SD 0.03) and 0.90 (SD 0.00) for EfficientNet-B4 and ResNet-50, respectively. All measures were within the range of corresponding interrater metrics. Conclusions We objectively benchmarked the explainability power of dermatological diagnosis models through the use of expert-defined supporting characteristics for diagnosis. Acknowledgments This work was supported in part by the Danish Innovation Fund under Grant 0153-00154A. Conflict of Interest None declared.

Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations

Artificial intelligence (AI) systems have increasingly achieved expert-level performance in medical imaging applications. However, there is growing concern that such AI systems may reflect and amplify human bias, and reduce the quality of their performance in historically under-served populations such as female patients, Black patients, or patients of low socioeconomic status. Such biases are especially troubling in the context of underdiagnosis, whereby the AI algorithm would inaccurately label an individual with a disease as healthy, potentially delaying access to care. Here, we examine algorithmic underdiagnosis in chest X-ray pathology classification across three large chest X-ray datasets, as well as one multi-source dataset. We find that classifiers produced using state-of-the-art computer vision techniques consistently and selectively underdiagnosed under-served patient populations and that the underdiagnosis rate was higher for intersectional under-served subpopulations, for example, Hispanic female patients. Deployment of AI systems using medical imaging for disease diagnosis with such biases risks exacerbation of existing care biases and can potentially lead to unequal access to medical treatment, thereby raising ethical concerns for the use of these models in the clinic.

Avoiding a replication crisis in deep-learning-based bioimage analysis.

Deep learning algorithms are powerful tools to analyse, restore and transform bioimaging data, increasingly used in life sciences research. These approaches now outperform most other algorithms for a broad range of image analysis tasks. In particular, one of the promises of deep learning is the possibility to provide parameter-free, one-click data analysis achieving expert-level performances in a fraction of the time previously required. However, as with most new and upcoming technologies, the potential for inappropriate use is raising concerns among the biomedical research community. This perspective aims to provide a short overview of key concepts that we believe are important for researchers to consider when using deep learning for their microscopy studies. These comments are based on our own experience gained while optimising various deep learning tools for bioimage analysis and discussions with colleagues from both the developer and user community. In particular, we focus on describing how results obtained using deep learning can be validated and discuss what should, in our views, be considered when choosing a suitable tool. We also suggest what aspects of a deep learning analysis would need to be reported in publications to describe the use of such tools to guarantee that the work can be reproduced. We hope this perspective will foster further discussion between developers, image analysis specialists, users and journal editors to define adequate guidelines and ensure that this transformative technology is used appropriately.

Clinically Applicable Segmentation of Head and Neck Anatomy for Radiotherapy: Deep Learning Algorithm Development and Validation Study.

BackgroundOver half a million individuals are diagnosed with head and neck cancer each year globally. Radiotherapy is an important curative treatment for this disease, but it requires manual time to delineate radiosensitive organs at risk. This planning process can delay treatment while also introducing interoperator variability, resulting in downstream radiation dose differences. Although auto-segmentation algorithms offer a potentially time-saving solution, the challenges in defining, quantifying, and achieving expert performance remain.ObjectiveAdopting a deep learning approach, we aim to demonstrate a 3D U-Net architecture that achieves expert-level performance in delineating 21 distinct head and neck organs at risk commonly segmented in clinical practice.MethodsThe model was trained on a data set of 663 deidentified computed tomography scans acquired in routine clinical practice and with both segmentations taken from clinical practice and segmentations created by experienced radiographers as part of this research, all in accordance with consensus organ at risk definitions.ResultsWe demonstrated the model’s clinical applicability by assessing its performance on a test set of 21 computed tomography scans from clinical practice, each with 21 organs at risk segmented by 2 independent experts. We also introduced surface Dice similarity coefficient, a new metric for the comparison of organ delineation, to quantify the deviation between organ at risk surface contours rather than volumes, better reflecting the clinical task of correcting errors in automated organ segmentations. The model’s generalizability was then demonstrated on 2 distinct open-source data sets, reflecting different centers and countries to model training.ConclusionsDeep learning is an effective and clinically applicable technique for the segmentation of the head and neck anatomy for radiotherapy. With appropriate validation studies and regulatory approvals, this system could improve the efficiency, consistency, and safety of radiotherapy pathways.

Abstract IA-03: Unsupervised resolution of intra- and inter-tumoral heterogeneity using deep learning

Abstract Applications of deep learning to histopathology has demonstrated expert-level performance, but approaches have been largely concentrated on supervised classification tasks requiring context-specific training and deployment. We demonstrate that deep neural networks, previously trained to recognize diverse histomorphologies in one environment can effectively extrapolate their learned representations to resolve clinically relevant intra- and inter-patient tissue pattern differences in other cancer types without explicit instruction or additional optimization. Moreover, this unsupervised approach highlighted numerous human misclassifications, suggesting that repositioning of existing histology-educated networks could provide scalable approaches for quality assurance and discovery of unappreciated subgroups of disease. Citation Format: Phedias Diamandis. Unsupervised resolution of intra- and inter-tumoral heterogeneity using deep learning [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr IA-03.

MVnet: automated time-resolved tracking of the mitral valve plane in CMR long-axis cine images with residual neural networks: a multi-center, multi-vendor study

BackgroundMitral annular plane systolic excursion (MAPSE) and left ventricular (LV) early diastolic velocity (e’) are key metrics of systolic and diastolic function, but not often measured by cardiovascular magnetic resonance (CMR). Its derivation is possible with manual, precise annotation of the mitral valve (MV) insertion points along the cardiac cycle in both two and four-chamber long-axis cines, but this process is highly time-consuming, laborious, and prone to errors. A fully automated, consistent, fast, and accurate method for MV plane tracking is lacking. In this study, we propose MVnet, a deep learning approach for MV point localization and tracking capable of deriving such clinical metrics comparable to human expert-level performance, and validated it in a multi-vendor, multi-center clinical population.MethodsThe proposed pipeline first performs a coarse MV point annotation in a given cine accurately enough to apply an automated linear transformation task, which standardizes the size, cropping, resolution, and heart orientation, and second, tracks the MV points with high accuracy. The model was trained and evaluated on 38,854 cine images from 703 patients with diverse cardiovascular conditions, scanned on equipment from 3 main vendors, 16 centers, and 7 countries, and manually annotated by 10 observers. Agreement was assessed by the intra-class correlation coefficient (ICC) for both clinical metrics and by the distance error in the MV plane displacement. For inter-observer variability analysis, an additional pair of observers performed manual annotations in a randomly chosen set of 50 patients.ResultsMVnet achieved a fast segmentation (<1 s/cine) with excellent ICCs of 0.94 (MAPSE) and 0.93 (LV e’) and a MV plane tracking error of −0.10 ± 0.97 mm. In a similar manner, the inter-observer variability analysis yielded ICCs of 0.95 and 0.89 and a tracking error of −0.15 ± 1.18 mm, respectively.ConclusionA dual-stage deep learning approach for automated annotation of MV points for systolic and diastolic evaluation in CMR long-axis cine images was developed. The method is able to carefully track these points with high accuracy and in a timely manner. This will improve the feasibility of CMR methods which rely on valve tracking and increase their utility in a clinical setting.

A patient-centric dataset of images and metadata for identifying melanomas using clinical context

Prior skin image datasets have not addressed patient-level information obtained from multiple skin lesions from the same patient. Though artificial intelligence classification algorithms have achieved expert-level performance in controlled studies examining single images, in practice dermatologists base their judgment holistically from multiple lesions on the same patient. The 2020 SIIM-ISIC Melanoma Classification challenge dataset described herein was constructed to address this discrepancy between prior challenges and clinical practice, providing for each image in the dataset an identifier allowing lesions from the same patient to be mapped to one another. This patient-level contextual information is frequently used by clinicians to diagnose melanoma and is especially useful in ruling out false positives in patients with many atypical nevi. The dataset represents 2,056 patients (20.8% with at least one melanoma, 79.2% with zero melanomas) from three continents with an average of 16 lesions per patient, consisting of 33,126 dermoscopic images and 584 (1.8%) histopathologically confirmed melanomas compared with benign melanoma mimickers.

Unsupervised Resolution of Histomorphologic Heterogeneity in Renal Cell Carcinoma Using a Brain Tumor-Educated Neural Network.

Applications of deep learning to histopathology have proven capable of expert-level performance, but approaches have largely focused on supervised classification tasks requiring context-specific training and deployment. More generalizable workflows that can be easily shared across subspecialties could help accelerate and broaden adoption. Here, we hypothesized that histology-optimized feature representations, generated by a convolutional neural network (CNN) during supervised learning, are transferable and can resolve meaningful differences in large-scale, discovery-type unsupervised analyses. We used a CNN, previously trained to recognize brain tumor histomorphologies, to extract 512 feature representations from > 550 digital whole-slide images (WSIs) of renal cell carcinomas (RCCs) from The Cancer Genome Atlas and other previously unencountered tumors. We use these extracted feature vectors to conduct unsupervised image-set clustering and analyze the clinical and biologic relevance of the intra- and interpatient subgroups generated. Within individual WSIs, feature-based clustering could reliably segment tumor regions and other relevant histopathologic subpatterns (eg, adenosquamous and poorly differentiated regions). Across the larger RCC cohorts, clustering extracted features generated subgroups enriched for clinically relevant subtypes (eg, papillary RCC) and outcomes (eg, survival). Importantly, individual feature activation mapping highlighted salient subtype-specific patterns and features of malignancies (eg, nuclear grade, sarcomatous change) contributing to subgroupings. Moreover, some proposed clusters were enriched for recurring, human-based RCC-subtype misclassifications. Our data support that CNNs, pretrained on large histologic datasets, can extend learned representations to novel scenarios and resolve clinically relevant intra- and interpatient tissue-pattern differences without explicit instruction or additional optimization. Repositioning of existing histology-educated networks could provide scalable approaches for image classification, quality assurance, and discovery of unappreciated patterns and subgroups of disease.

Deep active learning for interictal ictal injury continuum EEG patterns

ObjectivesSeizures and seizure-like electroencephalography (EEG) patterns, collectively referred to as “ictal interictal injury continuum” (IIIC) patterns, are commonly encountered in critically ill patients. Automated detection is important for patient care and to enable research. However, training accurate detectors requires a large labeled dataset. Active Learning (AL) may help select informative examples to label, but the optimal AL approach remains unclear. MethodsWe assembled >200,000 h of EEG from 1,454 hospitalized patients. From these, we collected 9,808 labeled and 120,000 unlabeled 10-second EEG segments. Labels included 6 IIIC patterns. In each AL iteration, a Dense-Net Convolutional Neural Network (CNN) learned vector representations for EEG segments using available labels, which were used to create a 2D embedding map. Nearest-neighbor label spreading within the embedding map was used to create additional pseudo-labeled data. A second Dense-Net was trained using real- and pseudo-labels. We evaluated several strategies for selecting candidate points for experts to label next. Finally, we compared two methods for class balancing within queries: standard balanced-based querying (SBBQ), and high confidence spread-based balanced querying (HCSBBQ). ResultsOur results show: 1) Label spreading increased convergence speed for AL. 2) All query criteria produced similar results to random sampling. 3) HCSBBQ query balancing performed best. Using label spreading and HCSBBQ query balancing, we were able to train models approaching expert-level performance across all pattern categories after obtaining ∼7000 expert labels. ConclusionOur results provide guidance regarding the use of AL to efficiently label large EEG datasets in critically ill patients.

Can artificial intelligence achieve human-level performance? A pilot study of childhood sexual abuse detection in self-figure drawings

Childhood sexual abuse (CSA) is a worldwide phenomenon that has negative long-term consequences for the victims and their families, and inflicts a considerable economic toll on society. One of the main difficulties in treating CSA is victims' reluctance to disclose their abuse, and the failure of professionals to detect it when there is no forensic evidence (Bottoms et al., 2014; McElvaney, 2013). Estimated disclosure rates for child sexual abuse based on retrospective adult reports range from 23 % to 45 % (e.g., Bottoms et al., 2014). This study reports the four stages in the development of a Convolutional Neural Network (CNN) system designed to detect abuse in self-figure drawings: (1) A preliminary study to build a Gender CNN; (2) Expert-level performance evaluation, (3) validation of the CSA CNN, (4) testing of the CSA CNN model. The findings indicate that the Gender CNN achieved 88 % detection accuracy and outperformed the CSA CNN by 19 percentage points. The CSA CNN achieved 72 % accuracy on the test set with 80 % precision and 79 % recall for the abuse class prediction. However, human experts outperformed the CSA CNN by 16 percentage points, probably due to the complexity of the task. These preliminary results suggest that CNN, when further developed, can contribute to the detection of child sexual abuse.

Becoming an expert: highly-experienced allied health professionals' relationships with their work

PurposeHighly-experienced allied health professionals have the opportunity to perform at the expert-level by sharing knowledge and skills with more junior staff, with the aim of upskilling the workforce. The study explored the current motivators, aspirations and the role of work in the life of highly-experienced practitioners, revealing factors that hinder or support them to further develop their own expertise and be inspiring role-models and mentors for less experienced staff.Design/methodology/approachTaking a grounded theoretical research design, we report on interviews with 45 allied health practitioners with at least seven years of professional experience from different professions and across organisational sectors. Transcripts were coded iteratively in conjunction with reviewing the literature, and cases were categorised to form a conceptual typology of work orientation.FindingsFour work orientations are characterised capturing the diversity of the allied health workforce, particularly in relation to two dimensions of personal fulfilment and future ambition. The relationship between the types and expert-level performance is discussed, leading to recommendations for support that can be implemented to develop and sustain expert-level performance within a community.Originality/valueA new view of work orientation is introduced that relates expert-level performance with meaningful work. The findings highlight a diversity of work orientation for highly-experienced allied health professionals, that all require managerial awareness. Once recognised, the four types would benefit from different supports that could develop and maintain expert-level performance in those that seek it. Alternatives are also available for those that do not. Implications for workforce policy are discussed.