Medical Imaging AI for Radiology

The Challenge

A leading regional healthcare network with 12 hospitals and over 200 radiologists faced a critical bottleneck in their diagnostic imaging workflow. With chest CT scan volumes increasing by 15% year-over-year and a nationwide shortage of radiologists, the organization struggled to maintain acceptable turnaround times while ensuring diagnostic accuracy.

The existing workflow required radiologists to manually review each slice of a CT scan, a process that could take 15-20 minutes per study. Lung nodule detection, crucial for early cancer diagnosis, was particularly time-consuming due to the subtle nature of early-stage nodules and the sheer volume of images. Fatigue-related errors were becoming a concern, and report turnaround times had stretched to 48+ hours for routine cases.

The Solution

Jasnova partnered with the healthcare network to develop and deploy a custom deep learning system specifically designed for lung nodule detection and characterization. Our approach combined cutting-edge computer vision technology with a deep understanding of radiologist workflows.

Technical Architecture

• 3D Convolutional Neural Network: Built a custom 3D CNN architecture optimized for volumetric CT analysis, capable of processing entire scan volumes rather than individual slices
• Multi-scale Detection: Implemented a feature pyramid network to detect nodules across a wide range of sizes, from 3mm micro-nodules to larger masses
• Attention Mechanisms: Integrated transformer-based attention to focus on clinically relevant regions while suppressing false positives from blood vessels and normal anatomical structures
• PACS Integration: Developed seamless integration with the existing PACS system through DICOM protocols, enabling automatic processing of incoming studies

Training and Validation

The model was trained on a curated dataset of over 50,000 annotated CT scans, with ground truth labels validated by board-certified thoracic radiologists. We employed extensive data augmentation techniques including elastic deformations, intensity variations, and synthetic nodule insertion to improve model robustness.

Rigorous validation was performed using a held-out test set of 5,000 scans, with performance benchmarked against both radiologist reads and existing CAD solutions. The final model achieved 95% sensitivity at a false positive rate of 2.1 per scan, significantly outperforming legacy CAD systems.

Implementation

The deployment was executed in three phases over 12 weeks:

Phase 1: Infrastructure Setup (Weeks 1-3)

Deployed on-premise GPU servers within the hospital's existing data center to ensure HIPAA compliance and minimize latency. Established secure connections to PACS systems across all 12 facilities.

Phase 2: Shadow Mode (Weeks 4-8)

The AI system processed all incoming chest CTs in shadow mode, with results reviewed by our clinical team and compared against radiologist reports. This phase allowed us to fine-tune detection thresholds and optimize the user interface based on radiologist feedback.

Phase 3: Production Rollout (Weeks 9-12)

Gradual rollout across facilities, beginning with high-volume sites. AI-generated findings were integrated directly into the radiologist's reading workflow, appearing as a structured checklist with annotated images for each detected finding.

Results

After six months of production use, the impact was substantial:

• 40% Reduction in Read Times: Average time per chest CT study decreased from 18 minutes to 11 minutes, as radiologists could quickly confirm or dismiss AI-flagged findings rather than exhaustively searching each slice
• 95% Detection Sensitivity: The system maintained 95% sensitivity for nodules 4mm and larger, with radiologists reporting increased confidence in their reads
• 23% Increase in Incidental Finding Detection: The AI consistently flagged small nodules that might have been overlooked during busy periods, leading to earlier detection of potentially malignant lesions
• Report Turnaround Improvement: Average turnaround time for routine chest CTs decreased from 48 hours to 18 hours
• Radiologist Satisfaction: 92% of radiologists reported that the tool improved their workflow and reduced cognitive burden

Key Learnings

This project reinforced several critical principles for deploying AI in clinical settings:

• Workflow Integration is Critical: The best AI model is useless if it disrupts existing workflows. Our tight PACS integration and intuitive interface were essential to adoption.
• Trust Requires Transparency: Radiologists needed to understand why the AI made specific detections. Our explainable AI features, including attention heatmaps and confidence scores, were crucial for building trust.
• Continuous Monitoring Matters: We implemented automated performance monitoring to detect any drift in model accuracy, ensuring consistent performance over time.

Technology Stack