Fully automated segmentation of foot bones using machine learning and convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Automated segmentation of foot bones is crucial for improving diagnostic accuracy and efficiency in medical imaging.
• A lack of literature on the reliability and validation of automated systems necessitates further research for clinical trust.

Purpose of the Study:

• To develop and implement a fully automated foot bone segmentation method using convolutional neural networks (CNNs).

Main Methods:

• Manual segmentation of foot bones from 50 CT scans.
• Training three-dimensional (3D) U-Net CNNs on 48 scans.
• Validation using Dice coefficient and Intersection over Union (IoU) on 2 test scans.

Main Results:

• The CNN achieved an overall Dice coefficient of 0.97 and IoU of 0.94.
• Accurate segmentation of hindfoot, midfoot, hallux, sesamoids, and proximal phalanges.
• Slightly lower performance for intermediate and distal phalanges of lesser toes.

Conclusions:

• CNN networks show excellent ability in recognizing foot bone structures on CT scans.
• Deep learning models show potential for reliable and accurate foot bone segmentation in clinical settings.