Deep learning-based breast lesion classification using treatment response assessment maps with uncertainty
Jerry Wang1, Bowen Jing2, Baowei Fei3
1University of Texas at Austin, Department of Biomedical Engineering, Austin, Texas, United States.
Purpose:
Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is widely used for breast cancer diagnosis due to its high sensitivity; however, limited specificity often leads to false positives and unnecessary biopsies. Treatment response assessment maps (TRAMs), derived from delayed contrast-enhanced imaging, characterize contrast accumulation and clearance and may improve diagnostic specificity.
Approach:
We utilized the Advanced MRI Breast Lesions dataset from The Cancer Imaging Archive to develop deep learning-based models for breast lesion classification using both multiphase DCE-MRI and TRAM images. A 3D convolutional residual network (ResNet18) was employed for feature extraction, followed by a fully connected classifier. To assess prediction reliability, test-time augmentation was used to estimate aleatoric uncertainty. Radiomics-based models were also implemented for comparison.
Results:
The TRAM-based deep learning model achieved numerically higher area under the receiver operating characteristic curve (0.870 versus 0.835), sensitivity (0.848 versus 0.818), and specificity (0.823 versus 0.759) compared with DCE-MRI. Radiomics analysis showed similar trends. Uncertainty stratification demonstrated improved performance in low-uncertainty predictions for both modalities, where high-uncertainty DCE-MRI predictions showed a marked performance decline compared with TRAM.
Conclusions:
These results suggest that TRAM-based deep learning models may provide a promising complementary imaging approach for breast lesion classification. Integration of uncertainty quantification could further enhance reliability, showing the potential of this framework for improving breast cancer diagnosis.