Project Overview
A Japanese medical device company specializing in diagnostic imaging equipment approached BCT Global with an ambitious goal: develop an AI-powered assistant that could pre-screen chest X-rays for anomalies, helping radiologists prioritize urgent cases. With Japan's aging population driving a 30% increase in diagnostic scan volumes over the past five years, radiologists were increasingly overwhelmed. The challenge was not just technical — the solution needed to meet stringent PMDA (Pharmaceuticals and Medical Devices Agency) regulatory requirements and ensure complete APPI compliance for patient data handling.
The Challenge
- Radiologists reviewing 150+ scans daily with rising burnout, leading to 8-12% of urgent findings being delayed in the queue
- The AI system needed to achieve clinical-grade accuracy (>90% sensitivity) to be useful as a pre-screening tool
- PMDA regulatory approval process required extensive documentation, validation protocols, and clinical testing data
- Patient data handling must comply with both APPI and medical-specific privacy regulations — all training data needed to be fully anonymized
- Integration with existing PACS (Picture Archiving and Communication System) via DICOM standard was essential — hospitals would not adopt a system that required workflow changes
- The client had no in-house AI/ML expertise and needed a team that could handle both the technical development and regulatory documentation
Our Solution
BCT Global deployed a specialized lab-style team of 6: 2 ML engineers with medical imaging experience, 1 backend developer, 1 DICOM/PACS integration specialist, 1 data engineer, and 1 project manager who coordinated with the client's regulatory affairs team.
- Phase 1 — Data Preparation & Model Research (Months 1-3): Worked with the client's partner hospitals to collect and anonymize 50,000+ chest X-ray images. Implemented a robust data pipeline with strict anonymization protocols. Evaluated multiple model architectures (ResNet, DenseNet, EfficientNet) and selected an ensemble approach for maximum reliability.
- Phase 2 — Model Development & Training (Months 3-6): Trained the anomaly detection model using PyTorch with transfer learning from pre-trained medical imaging models. Implemented attention mechanisms to highlight suspicious regions on X-rays. Achieved 94% sensitivity and 91% specificity on the validation dataset.
- Phase 3 — Integration & Validation (Months 6-8): Built a FastAPI-based inference server deployed on Azure (chosen for healthcare compliance features). Developed DICOM-compliant integration that fits seamlessly into existing radiology workflows. Created a web-based review interface showing AI findings alongside original images.
- Phase 4 — Regulatory & Deployment (Months 8-10): Prepared comprehensive PMDA submission documentation including validation protocols, clinical accuracy reports, and risk analysis. Deployed in 3 partner hospitals for clinical validation. Achieved PMDA approval for use as a Class II medical device software.
System Architecture
Technology Stack
Results & Impact
The AI pre-screening system reduced average radiologist review time by 35% by automatically prioritizing scans with detected anomalies. With 94% sensitivity, the system caught anomalies that might have been delayed in manual queue prioritization. The solution received PMDA approval within the initial project timeline — a rare achievement attributed to BCT Global's thorough regulatory documentation. Three partner hospitals adopted the system in the first month, with plans for nationwide distribution.
The client has since contracted BCT Global for Phase 2: expanding the model to detect specific pathologies including pneumonia, pleural effusion, and lung nodules.
"Developing an AI system for medical use in Japan requires navigating complex regulatory waters. BCT Global's team not only built an accurate model but also prepared documentation that our regulatory affairs team described as 'the most thorough they had ever reviewed from an external partner.' Their understanding of both the technical and regulatory aspects of healthcare AI was exceptional."
Chief Medical Officer Japanese Medical Device Company — PMDA-regulated, Tokyo