Breaking barriers: noninvasive AI model for BRAF^V600E mutation identification

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• BRAF^V600E mutation is common in papillary thyroid carcinoma (PTC).
• Current BRAF^V600E mutation detection methods are invasive.
• Noninvasive methods for identifying BRAF^V600E mutations are needed.

Purpose of the Study:

• To develop a noninvasive artificial intelligence (AI) model for identifying BRAF^V600E mutations in PTC.
• To extract radiomic features and utilize deep transfer learning (DTL) from ultrasound images.
• To combine DTL and radiomics for improved diagnostic capability.

Main Methods:

• Regions of interest (ROIs) were annotated in ultrasound images.
• Radiomic and DTL features were extracted and combined into a DTLR model.
• LASSO regression was used for feature selection, and eight machine learning methods were employed for model construction.
• Performance was evaluated using AUC, accuracy, sensitivity, and specificity. Grad-CAM was used for visualization.

Main Results:

• The DTLR model, particularly with ResNet152, effectively identified BRAF^V600E mutations.
• The optimal DTLR model achieved an AUC of 0.833, accuracy of 80.6%, sensitivity of 76.2%, and specificity of 81.7% in the validation set.
• The DTLR model outperformed standalone DTL and radiomics models.
• Grad-CAM visualization confirmed the model's ability to identify relevant ultrasound image features.

Conclusions:

• The ResNet152-based DTLR model shows significant potential for noninvasively identifying BRAF^V600E mutations in PTC patients using ultrasound.
• Grad-CAM can aid in visually stratifying BRAF mutations.
• Further multi-center validation with larger datasets is recommended.