Improvement in breast lesion classification utilizing deep learning and treatment response assessment maps (TRAMs)

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is crucial for breast lesion diagnosis but suffers from low specificity.
• Low specificity in DCE-MRI leads to unnecessary invasive procedures like biopsies.
• Treatment Response Assessment Map (TRAM) offers a novel approach to characterize contrast dynamics in breast lesions.

Purpose of the Study:

• To develop and evaluate a deep learning model for breast lesion classification using TRAM.
• To compare the diagnostic performance of a TRAM-based model against a standard DCE-MRI model.
• To assess the potential of TRAM in improving the specificity of breast lesion diagnosis.

Main Methods:

• A deep learning model utilizing a 3D convolutional residual network (ResNet18) was developed to analyze TRAM images.
• Image representations were extracted using ResNet18 and fed into a fully connected classifier for lesion classification.
• The TRAM-based model's performance was benchmarked against a model trained on standard multi-phase DCE-MRI data.

Main Results:

• The TRAM-based deep learning model achieved a higher area under the receiver operating characteristic curve (AUROC) of 0.870 compared to 0.835 for the standard DCE-MRI model.
• The TRAM model demonstrated superior sensitivity (0.848 vs. 0.818) and specificity (0.823 vs. 0.759).
• These findings indicate a significant improvement in diagnostic accuracy using TRAM-based analysis.

Conclusions:

• Deep learning analysis of TRAM shows enhanced diagnostic performance for breast lesions compared to standard DCE-MRI.
• TRAM-based models hold promise for improving the specificity of breast lesion diagnosis, potentially reducing unnecessary biopsies.
• This approach may significantly aid clinical decision-making in breast cancer diagnosis and treatment planning.