Convolutional Neural Network Using a Breast MRI Tumor Dataset Can Predict Oncotype Dx Recurrence Score

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Oncotype DX is a genetic test for ER+/HER2- invasive breast carcinoma, providing a Recurrence Score (RS).
• The test is invasive and costly, prompting research into alternative prediction methods like radiomics.

Purpose of the Study:

• To investigate the feasibility of using a convolutional neural network (CNN) to predict Oncotype DX RS from breast MRI data.
• To assess the diagnostic accuracy of the CNN model in classifying breast cancer risk groups.

Main Methods:

• A retrospective study analyzed MRI data from 134 patients with ER+/HER2- invasive ductal carcinoma.
• Tumor 3D segmentation was performed, and a CNN model with four convolutional layers was trained and validated using 5-fold cross-validation.
• The CNN was used for both three-class (low, intermediate, high risk) and two-class (low vs. intermediate/high risk) predictions.

Main Results:

• The CNN achieved 81% accuracy in three-class prediction (AUC 0.92) and 84% accuracy in two-class prediction (AUC 0.92).
• Specificities were 90% (three-class) and 81% (two-class), while sensitivities were 60% (three-class) and 87% (two-class).

Conclusions:

• Deep CNN architecture is capable of being trained to predict Oncotype DX RS.
• Radiomics using CNNs presents a promising, potentially non-invasive approach for predicting breast cancer recurrence risk.