From Pixels to Practice: An AI-Enabled Clinical Decision Support System for Bedside PICC Assessment in the ICU

Technical Summary: From Pixels to Practice: An AI-Enabled Clinical Decision Support System for Bedside PICC Assessment in the ICU

Problem Statement
Accurate localization of Peripherally Inserted Central Catheter (PICC) tips on bedside chest X-rays (CXRs) is critical for patient safety but remains a significant diagnostic challenge, particularly for junior clinicians in Intensive Care Units (ICUs). While CXR is the gold standard for post-insertion verification, bedside imaging is often compromised by suboptimal conditions, including patient positioning constraints, monitoring artifacts (e.g., electrodes, wires), and complex thoracic pathologies. These factors lead to low contrast and high noise, making tip identification difficult. Furthermore, reliance on radiologists for formal reporting introduces diagnostic delays, and junior staff often lack the specialized expertise required for accurate interpretation. Existing AI solutions for central venous catheters often suffer from selection bias, poor generalizability across different medical centers, and a lack of clinically relevant anatomical context (e.g., providing pixel coordinates rather than rib-level localization).

Methodology
The study proposes a dual-model, instance segmentation system based on the YOLOv11 architecture to bridge the gap between algorithmic development and bedside application.

• Dual-Model Strategy: Instead of processing the catheter in isolation, the system employs two parallel models:
  1. PICC Segmentation Model: Detects and segments the catheter line.
  2. Rib Segmentation Model: Segments the rib cage to serve as anatomical landmarks.
     This cascade approach mimics clinical reasoning by mapping the catheter tip's spatial coordinates to specific rib levels or intercostal spaces (e.g., "right 5th rib"), rather than providing abstract pixel data.
• Data Construction:
  • Internal Dataset: 346 high-quality ICU bedside CXRs from Shanghai Changhai Hospital (March 2019–March 2022), partitioned into training (n=311) and validation (n=35) sets.
  • External Dataset: 100 images randomly sampled from the public RANZCR CLiP repository to test cross-center generalizability.
  • Preprocessing: Contrast Limited Adaptive Histogram Equalization (CLAHE) with a clip limit of 6.0 was applied to enhance local contrast and suppress noise.
  • Annotation: Ground truth was established via independent, blinded annotation by two senior radiologists, with a third expert adjudicating discrepancies.
• Model Training: The YOLOv11l-seg model was initialized with ImageNet pre-trained weights. Training utilized a multi-task loss function (classification, bounding box regression, and segmentation mask loss) over 200 epochs with data augmentation (rotation, translation, noise injection) to improve robustness.
• Clinical Deployment: The model was encapsulated into a cross-platform tool named "TubeScan" (accessible via WeChat Mini Program and Web), enabling a "capture-upload-diagnose" workflow.
• Evaluation: A reader study was conducted with 14 junior ICU clinicians (≤2 years experience) using 10 challenging CXRs. Performance was measured across three modes: Manual (unassisted), AI-only, and AI-assisted. Metrics included diagnostic accuracy and interpretation time.

Key Contributions

1. Anatomically Referenced Localization: Unlike traditional segmentation models that output pixel masks, this system integrates rib segmentation to provide clinically actionable localization (rib-level), aligning with standard medical reporting protocols.
2. Robust Instance Segmentation: The study demonstrates that the YOLOv11 instance segmentation architecture outperforms traditional semantic segmentation models (U-Net, DeepLabV3+, FCN, etc.) in handling the severe class imbalance and artifact interference typical of ICU bedside CXRs.
3. Translational Framework ("From Pixels to Practice"): The work moves beyond theoretical model validation by deploying a functional, real-time decision support tool ("TubeScan") and rigorously testing its impact on clinical workflow and junior clinician performance.

Results

• Segmentation Performance:
  • Internal Validation: The PICC model achieved a Dice Similarity Coefficient (DSC) of 0.912, with Precision and Recall of 0.910 and 0.914, respectively. The rib model achieved an F1-score of 0.824.
  • External Validation: The PICC model maintained high performance with a DSC of 0.930 (Precision: 0.969, Recall: 0.892).
  • Comparison: The YOLO-based model significantly outperformed benchmark models. Notably, while the YOLO model maintained stability, benchmark models like DeepLabV3+ and U-Net showed marked performance degradation (e.g., DeepLabV3+ F1-score dropped from 0.66 to 0.53) during external validation, highlighting the YOLO model's superior generalizability.
• Clinical Utility (Reader Study):
  • AI Alone: Achieved 90.0% accuracy with a mean analysis time of 4.31 seconds.
  • Junior Clinicians (Manual): Achieved a mean accuracy of only 13.3% with a mean interpretation time of 20.8 seconds.
  • AI-Assisted: When junior clinicians used the "TubeScan" tool, their diagnostic accuracy improved approximately 7-fold to 95.0% (p<0.001), and interpretation time was reduced by more than half to 9.72 seconds (p<0.001).

Significance and Claims
The paper claims to provide a validated framework for translating AI from algorithmic research to practical clinical application in critical care settings. By addressing the specific challenges of ICU bedside imaging (artifacts, low quality) and the need for anatomical context, the system significantly reduces diagnostic uncertainty and time for junior medical staff. The authors position this work as an initial step toward multi-parameter machine learning in critical care, suggesting the potential adaptability of the model for other catheters (e.g., ECMO cannulas) and integration with other modalities. The study emphasizes that while the model achieves high precision, it serves as a decision support tool where final verification by clinicians remains necessary. The successful deployment of "TubeScan" demonstrates the feasibility of moving AI "from pixels to practice" to enhance precision medicine at the point of care.