Improved Segmentation and Detection Sensitivity of Diffusion-weighted Stroke Lesions with Synthetically Enhanced Deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Deep learning models require extensive, accurately labeled datasets for optimal performance in medical image analysis.
• Training deep learning models on limited clinical data can lead to suboptimal segmentation and detection of stroke lesions on diffusion-weighted (DW) images.
• Synthetic data generation offers a potential solution to augment limited clinical datasets.

Purpose of the Study:

• To compare the segmentation and detection performance of a deep learning model trained on human-labeled clinical stroke lesions versus one trained on a dataset augmented with synthetic stroke lesions.
• To evaluate the impact of synthetic data volume on the performance of deep learning models for stroke lesion analysis.
• To assess the sensitivity and specificity of the deep learning models against human expert performance.

Main Methods:

• A three-dimensional (3D) U-Net model was trained on four distinct datasets: clinical data only (CDB), synthetic data only (S2DB), CDB plus 2000 synthetic cases (CS2DB), and CDB plus 40,000 synthetic cases (CS40DB).
• Synthetic stroke lesions were generated by warping real stroke features onto normal brain volumes.
• Model performance was evaluated using Dice scores for segmentation accuracy and compared against three neuroradiologists for detection sensitivity and specificity.

Main Results:

• The model trained on the largest synthetic dataset (CS40DB) achieved the highest segmentation accuracy (Dice score, 0.72), outperforming models trained on less data (CS2DB: 0.70; CDB: 0.65).
• The CS40DB model demonstrated superior detection sensitivity (91%) compared to human readers (78%-84%).
• However, the CS40DB model exhibited lower specificity (75%) than human readers (89%-96%).

Conclusions:

• Augmenting deep learning training datasets with synthetic stroke lesions significantly enhances segmentation and detection performance on DW images.
• The volume of synthetic data is crucial, with larger datasets yielding better results.
• While synthetic data improves sensitivity, further refinement may be needed to match human reader specificity.