Computer Vision Development in New York

Healthcare / Hospital and Radiology

A New York City hospital system (18 hospitals, 6 million patient encounters per year, 280 radiologists processing 4.8 million diagnostic images annually) needed computer vision AI to address radiologist capacity constraints, reduce report turnaround times, and improve diagnostic accuracy for critical findings. (1) Radiologist capacity: the system's radiologists processed an average of 65 studies per day — approaching the upper limit recommended by the ACR for maintaining quality. During overnight and weekend shifts, 4 radiologists covered the entire 18-hospital system — creating critical bottleneck for emergency imaging. Turnaround time for emergency studies: 45 minutes average, with 12% exceeding the 60-minute target. (2) Critical finding detection: the most impactful diagnostic challenge was detecting critical findings (pneumothorax, pulmonary embolism, intracranial haemorrhage, and cervical spine fractures) on emergency studies. These conditions require immediate attention — a missed finding or delayed detection can be fatal. Despite high radiologist skill, the statistical reality of reviewing 65+ studies per day meant that critical findings were occasionally deprioritised in the reading queue simply because the worklist was ordered by arrival time rather than clinical urgency. (3) Measurement consistency: radiology reports required precise measurements — tumour sizes, cardiac dimensions, and anatomical measurements that informed treatment decisions. Inter-reader variability was 15-22% for common measurements — different radiologists measuring the same structure would report different values, creating confusion for referring clinicians and complicating longitudinal tracking of disease progression. (4) Follow-up tracking: 18% of radiology reports recommended follow-up imaging — but follow-up compliance was only 62%. Patients with concerning findings were lost to follow-up because the recommendation was buried in a text report that clinicians might not fully read, or because scheduling and patient communication systems were not integrated with radiology recommendations.

We developed and deployed a medical imaging AI platform over 16 weeks. (1) Critical finding prioritisation: deep learning models for detecting critical findings on emergency imaging — pneumothorax on chest X-ray (sensitivity 96.2%, specificity 94.8%), intracranial haemorrhage on head CT (sensitivity 97.1%, specificity 93.5%), pulmonary embolism on CT pulmonary angiogram (sensitivity 95.8%, specificity 92.1%), and cervical spine fracture on CT (sensitivity 94.5%, specificity 95.2%). Worklist integration: studies flagged as potentially containing critical findings were automatically elevated to the top of the radiologist's worklist — ensuring that the most urgent studies were read first regardless of arrival time. Alert workflow: when the AI detected a suspected critical finding with high confidence, an immediate alert was sent to the on-call radiologist and the referring emergency physician — enabling clinical action even before the formal radiology report was completed. (2) Automated measurements: computer vision models providing automated measurements for common radiology tasks — cardiac measurements on chest X-ray (cardiothoracic ratio — automated from PA chest X-ray with 2% measurement error vs. 15% inter-reader variability), pulmonary nodule sizing (automated longest diameter measurement on CT — enabling consistent longitudinal tracking for lung cancer screening), and fracture characterisation (displacement measurement, angulation, and fragment identification — providing structured measurements that reduced inter-reader variability from 22% to 5%). Measurement overlay: AI-generated measurements displayed alongside the images for radiologist review — the radiologist confirmed or adjusted the measurements rather than measuring from scratch. (3) Report assistance: structured finding generation — AI pre-populated structured findings based on image analysis. For chest X-ray, the system generated preliminary findings (lung opacity, pleural effusion, cardiomegaly, pneumothorax, rib fractures, and mediastinal abnormalities) with confidence scores. The radiologist reviewed, edited, and signed the report — reducing report generation time from 8 minutes to 4.5 minutes for chest X-rays. Natural language: findings presented in standard radiology language consistent with the department's reporting conventions — not requiring the radiologist to translate AI output into clinical language. (4) Follow-up tracking: NLP analysis of radiology reports identifying follow-up recommendations — extracting the recommended imaging modality, timeframe, and clinical reason. Patient tracking: automated monitoring of whether recommended follow-up imaging was scheduled and completed. Alert system: notifications to referring clinicians and care coordinators when follow-up imaging was overdue — with escalation workflows for high-priority findings (suspected malignancy, aneurysm monitoring). Patient communication: automated patient letters (in English, Spanish, and Chinese — the system's primary patient languages) reminding patients of recommended follow-up imaging with scheduling information.

Critical finding detection: average time-to-radiologist for critical findings reduced from 45 minutes to 12 minutes — AI prioritisation ensuring critical studies read first. The 60-minute target compliance improved from 88% to 99.2%. During overnight shifts, critical finding detection time improved by 65% — the AI flagging system compensated for the reduced overnight radiologist staffing. Zero delayed critical finding diagnoses attributable to worklist ordering in the 12 months following deployment (from 23 incidents in the prior year). Measurements: inter-reader variability for AI-assisted measurements reduced from 15-22% to 3-5% — dramatically improving consistency for longitudinal tracking. Pulmonary nodule measurement consistency enabled reliable growth assessment, reducing unnecessary biopsies by 18% (better measurement accuracy distinguished stable from growing nodules). Reporting efficiency: chest X-ray report time from 8 minutes to 4.5 minutes (44% reduction). Across 4.8 million annual studies, AI assistance saved approximately 520,000 radiologist-hours annually — equivalent to the capacity of 28 additional radiologists at current workload levels. Follow-up compliance: from 62% to 84% — automated tracking and patient communication reducing the number of patients lost to follow-up. For suspected malignancy findings, follow-up compliance reached 96% (from 75%). Estimated lives impacted: 340 patients per year received earlier diagnosis of significant conditions through improved follow-up compliance. Annual value: $54M (radiologist capacity — equivalent of 28 additional radiologists at $350K average fully loaded cost, offset by hiring 6 to handle growth), $18M (clinical outcomes — reduced delayed diagnoses, improved follow-up compliance, and avoided malpractice exposure), $12M (operational efficiency — faster report turnaround enabling better ED throughput and reduced patient wait times), $8M (measurement consistency — reduced unnecessary procedures from improved longitudinal tracking) = $92M. Development cost: $1.2M. Payback: under 5 days.