Semi-supervised Learning for Generalizable Intracranial Hemorrhage Detection and Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Intracranial hemorrhage (ICH) detection and segmentation are critical for diagnosing traumatic brain injuries.
• Deep learning models show promise but often require large labeled datasets, limiting their generalizability.
• Evaluating model performance on out-of-distribution datasets is crucial for real-world applicability.

Purpose of the Study:

• To develop and evaluate a semi-supervised learning (SSL) model for ICH detection and segmentation.
• To assess the model's generalizability on an out-of-distribution head CT dataset.
• To compare the SSL model's performance against a traditional supervised baseline.

Main Methods:

• A "teacher" deep learning model was trained on 457 labeled head CT scans.
• Pseudo-labels were generated on 25,000 unlabeled scans using the teacher model.
• A "student" model was trained on the combined labeled and pseudo-labeled data, with hyperparameter tuning on a validation set.
• Performance was evaluated on the CQ500 dataset for out-of-distribution generalizability.

Main Results:

• The SSL model achieved a statistically significant higher area under the receiver operating characteristic curve (0.939 vs 0.907, P = .009) on the CQ500 dataset.
• The SSL model demonstrated a higher Dice similarity coefficient (0.829 vs 0.809, P = .012) and pixel average precision (0.848 vs 0.828) compared to the baseline.
• These results indicate improved performance in both classification and segmentation tasks.

Conclusions:

• Incorporating unlabeled data via semi-supervised learning enhances the generalizability of ICH detection and segmentation models.
• SSL models show superior performance on out-of-distribution datasets compared to purely supervised approaches.
• This approach holds significant potential for improving AI-driven diagnostic tools in neuroradiology.