Building RadScan AI: Autonomous Multimodal Radiology Triage on GCP Cloud Run & Vertex AI

说明与用途

RadScan AI 是一个自主的放射科助手，利用 GCP Cloud Run、Vertex AI Gemini 和 CNN-BiGRU 神经网络分析 DICOM 扫描。该系统自动检测病理，创建 Grad-CAM 热图，并在 3 秒内起草报告，将放射科医生的时间从 10-15 分钟减少到 6 分钟。

问题

Radiologists face extreme burnout, evaluating 50+ volumetric MRI scans daily. Each scan contains 24 to 40 high-resolution DICOM slices. Manually inspecting every plane, checking ligament continuity across depth slices, and dictating repetitive clinical reports takes 10 to 15 minutes per study.

结果

RadScan AI reduces radiology report drafting time from 10-15 minutes to under 3 seconds per study, saving approximately 6 minutes per scan. It achieves this by combining a 2.5D Volumetric CNN-BiGRU engine for pathology detection and Grad-CAM heatmaps for visual explanation, with Vertex AI Gemini for structured report generation.

工作流

• Client (Next.js 14 Medical Workspace) uploads DICOM files via 1-Click Sample Buttons or upload interface.
• Frontend displays Multi-Planar Slice Slider and Grad-CAM Heatmap Opacity Layer.
• Backend Microservice (FastAPI on GCP Cloud Run L4 GPU) receives POST /api/v1/predict request.
• 2.5D Volumetric CNN-BiGRU Engine processes 24 parallel depth slices across Sagittal, Coronal, and Axial planes.
• Grad-CAM Heatmap Synthesizer generates visual heatmaps pinpointing lesion coordinates.
• Backend sends results to LLM Engine (Vertex AI Gemini 3.5 Flash) via POST /api/v1/report.
• Gemini generates structured DICOM reports in JSON format.
• Frontend displays the clinical report and interactive heatmap to the radiologist.

证据

Manually inspecting every plane, checking ligament continuity across depth slices, and dictating repetitive clinical reports takes 10 to 15 minutes per study.

来源

drafts structured DICOM reports in under 3 seconds—saving radiologists ~6 minutes per scan.

来源

trained on 819,100 DICOMs / 530 GB data

来源