Intelligent Recognition and Segmentation of Blunt Craniocerebral Injury CT Images Based on DeepLabV3+ Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Forensic Medicine

Background:

• Blunt craniocerebral injury (BCI) diagnosis relies on accurate image analysis.
• Automated segmentation of BCI on CT images can aid forensic pathology.
• Deep learning offers potential for automated medical image analysis.

Purpose of the Study:

• To develop and evaluate a DeepLabV3+ model for segmenting common BCI types from CT images.
• To explore the efficacy of a deep learning model trained on living individuals for cadaveric BCI identification.

Main Methods:

• Trained a DeepLabV3+ convolutional neural network on 5,486 CT images of BCI from living individuals.
• Validated the model on separate test sets of living and cadaveric BCI CT images.
• Evaluated segmentation performance using Dice, accuracy, precision, and F1 scores.

Main Results:

• The model achieved high Dice values (>0.75) for segmenting scalp hematoma, skull fracture, epidural hematoma, subdural hematoma, and brain contusion in living individuals.
• External validation showed strong F1 scores for most BCI types in living individuals.
• Segmentation performance was reduced but still notable for scalp hematoma, epidural hematoma, and subdural hematoma in cadaveric CT images.

Conclusions:

• Deep learning models trained on CT images are effective for BCI segmentation in living individuals.
• Direct application of models trained on living BCI to cadaveric BCI presents limitations.
• This study introduces a novel method for intelligent segmentation of BCI in virtual anatomical data.