Predicting malignancy in breast lesions: enhancing accuracy with fine-tuned convolutional neural network models

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate prediction of breast cancer malignancy is crucial for effective patient management.
• Dynamic Contrast-Enhanced Breast Magnetic Resonance Imaging (DCE-BMRI) is a key imaging modality for breast lesion characterization.
• Convolutional Neural Network (CNN) models offer potential for automated analysis of medical images.

Purpose of the Study:

• To evaluate the accuracy of various CNN models in distinguishing malignant from benign breast lesions on DCE-BMRI.
• To identify the optimal CNN architecture for predicting malignancy in breast MRI.
• To assess the performance of CNN models across different BI-RADS categories.

Main Methods:

• A dataset of 547 lesions (273 benign, 274 malignant) was split into training, testing, and validation sets.
• Five established CNN models (VGG16, VGG19, DenseNet201, ResNet50, MobileNetV2) were evaluated.
• A novel S4 model was developed and compared against the five standard models using metrics like accuracy, precision, recall, F1 score, and AUC.

Main Results:

• The VGG19 model achieved the highest accuracy (0.96) on the test set among the standard models.
• The proposed S4 model demonstrated superior performance on the validation set, with Precision (Pr) of 0.89, Recall (Rc) of 0.88, F1 score (F1) of 0.87, and Area Under the Curve (AUC) of 0.89.
• The S4 model showed significantly higher AUC for BI-RADS 3 (0.90) and BI-RADS 4 (0.86) lesions compared to BI-RADS 5 (0.65).

Conclusions:

• The S4 model exhibits excellent performance in predicting malignancy from DCE-BMRI, outperforming other evaluated CNNs.
• The S4 model shows promise for clinical application in breast disease diagnosis.
• Further external validation with larger datasets is recommended to confirm the S4 model's efficacy and generalizability.