Deep Learning for FAST Quality Assessment.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Ultrasound technology

Background:

• Focused assessment with sonography in trauma (FAST) exams are crucial for diagnosing injuries in emergency settings.
• Assessing the quality of FAST exams is vital for accurate diagnosis but can be subjective.
• Deep learning (DL) offers potential for automating and standardizing quality assessment.

Purpose of the Study:

• To evaluate the feasibility of a DL algorithm for assessing the quality of FAST exams.
• To compare the performance of different DL approaches for FAST exam quality assessment.

Main Methods:

• A dataset of 441 FAST exams (3161 videos) was used, classified as good or poor quality.
• Convolutional Neural Networks (CNNs) pretrained on Imagenet were fine-tuned.
• A CNN autoencoder was trained for image compression (20:1 ratio), with compressed codes feeding a classifier.

Main Results:

• Encoder-classifier networks significantly outperformed transfer learning with CNNs.
• The Imagenet dataset was found to be a poor match for ultrasound quality assessment.
• DL models achieved high performance with 99% sensitivity and 98% specificity on test sets.

Conclusions:

• Autoencoder-based image compression is highly effective for predicting FAST exam quality.
• Features derived from autoencoders are more suitable for FAST quality assessment than those from Imagenet-pretrained CNNs.