ABSTRACT
BACKGROUND: Artificial intelligence (AI) systems for automated chest x-ray interpretation hold promise for standardising reporting and reducing delays in health systems with shortages of trained radiologists. Yet, there are few freely accessible AI systems trained on large datasets for practitioners to use with their own data with a view to accelerating clinical deployment of AI systems in radiology. We aimed to contribute an AI system for comprehensive chest x-ray abnormality detection. METHODS: In this retrospective cohort study, we developed open-source neural networks, X-Raydar and X-Raydar-NLP, for classifying common chest x-ray findings from images and their free-text reports. Our networks were developed using data from six UK hospitals from three National Health Service (NHS) Trusts (University Hospitals Coventry and Warwickshire NHS Trust, University Hospitals Birmingham NHS Foundation Trust, and University Hospitals Leicester NHS Trust) collectively contributing 2 513 546 chest x-ray studies taken from a 13-year period (2006-19), which yielded 1 940 508 usable free-text radiological reports written by the contemporary assessing radiologist (collectively referred to as the "historic reporters") and 1 896 034 frontal images. Chest x-rays were labelled using a taxonomy of 37 findings by a custom-trained natural language processing (NLP) algorithm, X-Raydar-NLP, from the original free-text reports. X-Raydar-NLP was trained on 23 230 manually annotated reports and tested on 4551 reports from all hospitals. 1 694 921 labelled images from the training set and 89 238 from the validation set were then used to train a multi-label image classifier. Our algorithms were evaluated on three retrospective datasets: a set of exams sampled randomly from the full NHS dataset reported during clinical practice and annotated using NLP (n=103 328); a consensus set sampled from all six hospitals annotated by three expert radiologists (two independent annotators for each image and a third consultant to facilitate disagreement resolution) under research conditions (n=1427); and an independent dataset, MIMIC-CXR, consisting of NLP-annotated exams (n=252 374). FINDINGS: X-Raydar achieved a mean AUC of 0·919 (SD 0·039) on the auto-labelled set, 0·864 (0·102) on the consensus set, and 0·842 (0·074) on the MIMIC-CXR test, demonstrating similar performance to the historic clinical radiologist reporters, as assessed on the consensus set, for multiple clinically important findings, including pneumothorax, parenchymal opacification, and parenchymal mass or nodules. On the consensus set, X-Raydar outperformed historical reporter balanced accuracy with significance on 27 of 37 findings, was non-inferior on nine, and inferior on one finding, resulting in an average improvement of 13·3% (SD 13·1) to 0·763 (0·110), including a mean 5·6% (13·2) improvement in critical findings to 0·826 (0·119). INTERPRETATION: Our study shows that automated classification of chest x-rays under a comprehensive taxonomy can achieve performance levels similar to those of historical reporters and exhibit robust generalisation to external data. The open-sourced neural networks can serve as foundation models for further research and are freely available to the research community. FUNDING: Wellcome Trust.
Subject(s)
Artificial Intelligence , Image Interpretation, Computer-Assisted , Neural Networks, Computer , Humans , Retrospective Studies , X-RaysABSTRACT
ABSTRACT: A 43-year-old woman, who presented with a suspected left breast abscess, underwent serial ultrasounds, which demonstrated inflammatory changes that were nonresponsive to antibiotics and which spread to the contralateral breast. 18 F-FDG PET/CT demonstrated diffuse heterogeneous intense FDG uptake in both breasts with reactive axillary nodes. Breast biopsy confirmed granulomatous inflammation, and overall findings were consistent with idiopathic granulomatous mastitis. In the absence of histological analysis, idiopathic granulomatous mastitis is an important differential diagnosis to consider for bilateral abnormal breast uptake, and early recognition can facilitate prompt commencement of treatment.
Subject(s)
Fluorodeoxyglucose F18 , Granulomatous Mastitis , Female , Humans , Adult , Granulomatous Mastitis/diagnostic imaging , Positron Emission Tomography Computed Tomography , Positron-Emission Tomography , BreastABSTRACT
OBJECTIVES: A published tumour regression grade (TRG) score for squamous anal carcinoma treated with definitive chemoradiotherapy based on T2-weighted MRI yields a high proportion of indeterminate responses (TRG-3). We investigate whether the addition of diffusion-weighted imaging (DWI) improves tumour response assessment in the early post treatment period. MATERIALS AND METHODS: This retrospective observational study included squamous anal carcinoma patients undergoing MRI before and within 3 months of completing chemoradiotherapy from 2009 to 2020. Four independent radiologists (1-20 years' experience) scored MRI studies using a 5-point TRG system (1 = complete response; 5 = no response) based on T2-weighted sequences alone, and then after a 12-week washout period, using a 5-point DWI-TRG system based on T2-weighted and DWI. Scoring confidence was recorded on a 5-point scale (1 = low; 5 = high) for each reading and compared using the Wilcoxon test. Indeterminate scores (TRG-3) from each reading session were compared using the McNemar test. Interobserver agreement was assessed using kappa statistics. RESULTS: Eighty-five patients were included (mean age, 59 years ± 12 [SD]; 55 women). T2-weighted TRG-3 scores from all readers combined halved from 24% (82/340) to 12% (41/340) with DWI (p < 0.001). TRG-3 scores changed most frequently (41%, 34/82) to DWI-TRG-2 (excellent response). Complete tumour response was recorded clinically in 77/85 patients (91%). Scoring confidence increased using DWI (p < 0.001), with scores of 4 or 5 in 84% (287/340). Interobserver agreement remained fair to moderate (kappa range, 0.28-0.58). CONCLUSION: DWI complements T2-weighted MRI by reducing the number of indeterminate tumour responses (TRG-3). DWI increases radiologist's scoring confidence. CLINICAL RELEVANCE STATEMENT: Diffusion-weighted imaging improves T2-weighted tumour response assessment in squamous anal cancer, halving the number of indeterminate responses in the early post treatment period, and increases radiologists' confidence. KEY POINTS: Tumour response based on T2-weighted MRI is often indeterminate in squamous anal carcinoma. Diffusion-weighted imaging alongside T2-weighted MRI halved indeterminate tumour regression grade scores assigned by four radiologists from 24 to 12%. Scoring confidence of expert and non-expert radiologists increased with the inclusion of diffusion-weighted imaging.
Subject(s)
Anus Neoplasms , Carcinoma, Squamous Cell , Humans , Female , Middle Aged , Magnetic Resonance Imaging/methods , Diffusion Magnetic Resonance Imaging/methods , Anus Neoplasms/diagnostic imaging , Anus Neoplasms/therapy , Anus Neoplasms/pathology , Carcinoma, Squamous Cell/diagnostic imaging , Carcinoma, Squamous Cell/therapy , Carcinoma, Squamous Cell/pathology , Chemoradiotherapy , Retrospective StudiesABSTRACT
OBJECTIVE: A simulator to enable safe practice and assessment of ALND has been designed, and face, content and construct validity has been investigated. SUMMARY AND BACKGROUND DATA: The reduction in the number of ALNDs conducted has led to decreased resident exposure and confidence. METHODS: A cross-sectional multicenter observational study was carried out between July 2017 and August 2018. Following model development, 30 surgeons of varying experience (n = "experts," n = 11 "senior residents," and n = 10 "junior residents") were asked to perform a simulated ALND. Face and content validity questionnaires were administered immediately after ALND. All ALND procedures were retrospectively assessed by 2 attending breast surgeons, blinded to operator identity, using a video-based assessment tool, and an end product assessment tool. RESULTS: Statistically significant differences between groups were observed across all operative subphases on the axillary clearance assessment tool (P < 0.001). Significant differences between groups were observed for overall procedure quality (P < 0.05) and total number of lymph nodes harvested (P < 0.001). However, operator grade could not be distinguished across other end product variables such as axillary vein damage (P = 0.864) and long thoracic nerve injury (P = 0.094). Overall, participants indicated that the simulator has good anatomical (median score >7) and procedural realism (median score >7). CONCLUSIONS: Video-based analysis demonstrates construct validity for ALND assessment. Given reduced ALND exposure, this simulation is a useful adjunct for both technical skills training and formative Deanery or Faculty administered assessments.
Subject(s)
Clinical Competence , Lymph Node Excision/standards , Axilla , Cross-Sectional Studies , Humans , Retrospective StudiesABSTRACT
This study seeks to define long-term variation in body composition in patients undergoing esophagectomy for cancer and to associate those changes with survival. Assessment of skeletal muscle, visceral (VAT) and subcutaneous adipose tissue (SAT) was performed using computed tomography (CT) images routinely acquired: at diagnosis; after neoadjuvant therapy, and; >6 months after esophagectomy. In cases where multiple CT scans were performed >6 months after surgery, all available images were assessed. Ninty-seven patients met inclusion criteria with a median of 2 (range 1-10) postoperative CT images acquired between 0.5 and 9.7 years after surgery. Following surgical treatment of esophageal cancer, patients lost on average 13.3% of their skeletal muscle, 64.5% of their VAT and 44.2% of their SAT. Sarcopenia at diagnosis was not associated with worse overall survival (66.3% vs. 68.5%; P = 0.331). Sarcopenia 1 year after esophagectomy was however associated with lower 5-year overall survival (53.8% vs. 87.5%; P = 0.019). Survival was lower in those patients who had >10% decrease in skeletal muscle index (SMI; 33.3% vs. 72.1%; P = 0.003) and >40% decrease in SAT 1 year after surgery (40.4% vs. 67.4%; P = 0.015). On multivariate analysis, a decline in SMI 1 year after surgery was predictive of worse survival (HR 0.38, 95%CI 0.20-0.73; P = 0.004). This study provides new insight relating to long-term variation in body composition in patients undergoing esophagectomy for cancer. Findings provide further evidence of the importance of body composition, in particular depletion of skeletal muscle, in predicting survival following esophagectomy.
Subject(s)
Esophagectomy , Sarcopenia , Adiposity , Esophagectomy/adverse effects , Humans , Muscle, Skeletal/diagnostic imaging , Prognosis , Retrospective Studies , Sarcopenia/diagnostic imaging , Sarcopenia/etiologyABSTRACT
OBJECTIVES: To explore the attitudes of United Kingdom (UK) medical students regarding artificial intelligence (AI), their understanding, and career intention towards radiology. We also examine the state of education relating to AI amongst this cohort. METHODS: UK medical students were invited to complete an anonymous electronic survey consisting of Likert and dichotomous questions. RESULTS: Four hundred eighty-four responses were received from 19 UK medical schools. Eighty-eight percent of students believed that AI will play an important role in healthcare, and 49% reported they were less likely to consider a career in radiology due to AI. Eighty-nine percent of students believed that teaching in AI would be beneficial for their careers, and 78% agreed that students should receive training in AI as part of their medical degree. Only 45 students received any teaching on AI; none of the students received such teaching as part of their compulsory curriculum. Statistically, students that did receive teaching in AI were more likely to consider radiology (p = 0.01) and rated more positively to the questions relating to the perceived competence in the post-graduation use of AI (p = 0.01-0.04); despite this, a large proportion of students in the taught group reported a lack of confidence and understanding required for the critical use of healthcare AI tools. CONCLUSIONS: UK medical students understand the importance of AI and are keen to engage. Medical school training on AI should be expanded and improved. Realistic use cases and limitations of AI must be presented to students so they will not feel discouraged from pursuing radiology.
Subject(s)
Breast Neoplasms/surgery , Lymph Node Excision/education , Axilla , Female , Humans , Surveys and QuestionnairesABSTRACT
INTRODUCTION: Transaxillary robotic thyroid surgery was pioneered in South Korea where cultural factors, anthropometry and remuneration favour this. Small thyroid nodules account for the majority of cases due to a national thyroid cancer screening programme. However, the technique has not been evaluated in the United Kingdom where larger thyroid nodules tend to undergo surgery in a patient population with a higher body mass index (BMI). METHODS: Long term prospective non-randomised study. Sixteen consecutive robotic hemithyroidectomy patients were compared to 16 open controls. RESULTS: There were no robotic conversions to open and no significant difference regarding pain, voice, or quality of life (QoL). In the robotic group, long term. scar cosmesis at 3 years was superior (p = 0.02) although the operative time was significantly longer (228 min vs. 85 min, p = 0.01). One patient had a transient recurrent laryngeal nerve paresis and another had temporary shoulder dysfunction. Both resolved in 4 weeks. DISCUSSION: This study highlights the considerable difference between a Western patient population compared to South East Asia. Despite this robotic thyroidectomy is feasible and safe in a UK population. CONCLUSIONS: Despite a low uptake in the UK compared to the Far East, robotic thyroidectomy represents a viable option for selected patients, achieving superior cosmesis compared with conventional thyroidectomy at the expense of time and cost. The key is careful patient selection. A randomised study is needed to establish the clinical efficacy compared to conventional surgery in this population.
Subject(s)
Postoperative Complications/etiology , Robotic Surgical Procedures/adverse effects , Thyroid Nodule/surgery , Thyroidectomy/adverse effects , Adult , Aged , Body Mass Index , Case-Control Studies , Cicatrix , Cross-Cultural Comparison , Female , Humans , Male , Middle Aged , Operative Time , Prospective Studies , Quality of Life , Republic of Korea/ethnology , Robotic Surgical Procedures/methods , Thyroidectomy/methods , Treatment Outcome , United Kingdom/ethnologyABSTRACT
OBJECTIVES/HYPOTHESIS: To investigate the reliability and construct validity of procedure-based assessment (PBA) in assessing performance and progress in otolaryngology training. STUDY DESIGN: Retrospective database analysis using a national electronic database. METHODS: We analyzed PBAs of otolaryngology trainees in North London from core trainees (CTs) to specialty trainees (STs). The tool contains six multi-item domains: consent, planning, preparation, exposure/closure, technique, and postoperative care, rated as "satisfactory" or "development required," in addition to an overall performance rating (pS) of 1 to 4. Individual domain score, overall calculated score (cS), and number of "development-required" items were calculated for each PBA. Receiver operating characteristic analysis helped determine sensitivity and specificity. RESULTS: There were 3,152 otolaryngology PBAs from 46 otolaryngology trainees analyzed. PBA reliability was high (Cronbach's α 0.899), and sensitivity approached 99%. cS correlated positively with pS and level in training (rs : +0.681 and +0.324, respectively). ST had higher cS and pS than CT (93% ± 0.6 and 3.2 ± 0.03 vs. 71% ± 3.1 and 2.3 ± 0.08, respectively; P < .001). cS and pS increased from CT1 to ST8 showing construct validity (rs : +0.348 and +0.354, respectively; P < .001). The technical skill domain had the highest utilization (98% of PBAs) and was the best predictor of cS and pS (rs : +0.96 and +0.66, respectively). CONCLUSIONS: PBA is reliable and valid for assessing otolaryngology trainees' performance and progress at all levels. It is highly sensitive in identifying competent trainees. The tool is used in a formative and feedback capacity. The technical domain is the best predictor and should be given close attention. LEVEL OF EVIDENCE: NA.