Artificial Intelligence-Enabled Detection of Vascular Perfusion Defects on Ventilation/Perfusion (V/Q) Scintigraphy for Pulmonary Embolism

Technical Summary: Artificial Intelligence-Enabled Detection of Vascular Perfusion Defects on V/Q Scintigraphy

Problem Statement
The diagnosis of pulmonary embolism (PE) via planar ventilation/perfusion (V/Q) scintigraphy relies on the identification of mismatched V/Q defects. However, manual interpretation is cognitively demanding, time-consuming, and subject to significant interobserver variability. Current guidelines (PIOPED/EANM) require clinicians to visually estimate defect extent relative to bronchopulmonary segments, a process rarely standardized in practice due to the lack of objective automated tools. While AI has advanced in other nuclear medicine domains, V/Q planar imaging has seen a 17-year gap in modern deep learning applications for defect segmentation, with prior studies relying on shallow networks or radiomics rather than pixel-level segmentation.

Methodology
The study utilized a retrospective dataset of 2,118 patients from The Ottawa Hospital, comprising 9,852 planar V/Q projections across six standard views (ANT, POST, LAO, RAO, LPO, RPO). The data was split into training (1,313 patients), validation (329 patients), a high-probability hold-out test set (46 patients), and a normal hold-out set (430 patients).

• Annotation: Eight nuclear medicine physicians (fellows and a senior physician) annotated vascular perfusion mismatches. Ground truth for the hold-out set was established via an ensemble of seven annotators, where a pixel was considered a defect if segmented by ≥4 annotators (>50% consensus).
• Model Architecture: Four 2D deep learning models were trained to segment perfusion defects:
  1. U-Net: A baseline fully convolutional network.
  2. nnU-Net: A self-configuring U-Net variant.
  3. Swin UNETR: A U-Net-like architecture using a hierarchical Swin Transformer encoder.
  4. BTU-Net (Bottleneck Transformer U-Net): A novel hybrid architecture proposed by the authors. It integrates a standard U-Net contracting/expansive path with a transformer block in the bottleneck. This bottleneck applies multi-head self-attention (4 heads) to capture long-range dependencies before passing features to the expansive path, aiming to balance local detail with global context.
• Training: Models were trained using paired ventilation and perfusion images normalized to [0,1]. Loss functions combined Dice Similarity Coefficient (DSC) and binary cross-entropy. Training utilized the Adam optimizer with early stopping.
• Evaluation Metrics:
  • Segmentation Quality: For true positives, performance was measured via DSC, Jaccard Index, Mean Surface Distance (MSD), and Hausdorff Distance (HD/HD95).
  • Detection Accuracy: Free-response Receiver Operating Characteristic (FROC) analysis was used to evaluate sensitivity against the False Positive Per Projection Rate (FPPR).
  • Specificity: FPPR was calculated on the normal hold-out set (430 patients) to quantify spurious defect detection.

Key Results

• Segmentation Performance: The BTU-Net model consistently outperformed the other three architectures (U-Net, nnU-Net, Swin UNETR) across all segmentation metrics. It achieved a median DSC of 0.72 and a median Jaccard index of 0.57, with lower boundary error distances (MSD and HD95) compared to baselines.
• Clinical Detection (FROC): On the high-probability hold-out set, BTU-Net was the only model performing on par with human readers. At an FPPR of 1.5, BTU-Net achieved a sensitivity of 0.529 ± 0.026, compared to 0.446 for Swin UNETR and 0.355 for U-Net. Notably, human readers also exhibited significant variability, with sensitivities ranging from 0.40 to 0.60 depending on the observer, indicating that the AI model's performance aligns with the upper bound of human capability.
• False Positives: On the normal hold-out set (430 patients), BTU-Net demonstrated the lowest false positive rate (0.08 ± 0.01 FPPR), significantly outperforming Swin UNETR (0.12), nnU-Net (0.22), and U-Net (0.15) (P<0.0001).
• Qualitative Assessment: Visual inspection revealed that baseline U-Net often merged distinct defects and included normal tissue, while Swin UNETR produced imprecise shapes. BTU-Net generated more anatomically plausible, wedge-shaped defects consistent with lung segments.

Significance and Claims
The paper claims to be the first study to directly segment vascular perfusion defects on V/Q planar images using modern deep learning and to evaluate interobserver variability at the defect level. The authors assert that the BTU-Net architecture successfully revitalizes AI in V/Q imaging by achieving diagnostic accuracy comparable to expert nuclear medicine physicians.

The significance of this work lies in:

1. Standardization: Providing a tool to objectively delineate defects, potentially reducing interobserver variability and standardizing reporting.
2. Clinical Utility: Offering a decision-support tool that could assist non-expert readers, potentially increasing diagnostic confidence and reducing unnecessary referrals to CT pulmonary angiography (CTPA) for intermediate-probability cases.
3. Foundation for Future Work: Establishing a benchmark for V/Q mismatch segmentation that could be extended to other pathologies like chronic thromboembolic pulmonary hypertension (CTEPH) or used to train novice clinicians.

The authors maintain a modest tone, acknowledging that subsegmental defects remain challenging and that the model should serve as a "diagnostic aid with a human in the loop." They also note limitations regarding generalizability to other imaging protocols and the lack of a true gold standard (pulmonary angiography) for training, relying instead on the "second-best" standard of consensus human annotation.