A deep-learning semantic segmentation approach to fully automated MRI-based left-ventricular deformation analysis in...

Area of Science:

• Cardiovascular Imaging
• Medical Artificial Intelligence
• Oncology

Background:

• Left-ventricular (LV) strain analysis using Displacement Encoding with Stimulated Echoes (DENSE) MRI is crucial for estimating cardiotoxicity from breast cancer chemotherapy.
• Accurate quantification of LV parameters is essential for monitoring treatment-related cardiac damage.

Purpose of the Study:

• To investigate an automated LV chamber quantification tool using a supervised deep convolutional neural network (DCNN) for segmentation of DENSE MRI images.
• To evaluate the DCNN's performance in quantifying LV parameters and analyzing myocardial strain for cardiotoxicity assessment.

Main Methods:

• A custom DeepLabV3+ DCNN with a ResNet-50 backbone was trained and validated on 42 female breast cancer patient datasets.
• Segmentation accuracy was assessed using metrics including accuracy, Dice, and average perpendicular distance (APD).
• LV chamber quantification and myocardial strain results were validated against ground-truth, SSFP acquisitions, and a vendor tool-based method using Cronbach's Alpha (C-Alpha) correlation.

Main Results:

• The DCNN achieved high accuracy (97%), Dice score (0.89), and low APD (2.4 mm) for myocardial classification on the test set.
• High C-Alpha correlations (0.97 for LVEF, 0.77 for LVEDD) were observed between DENSE-based quantification and SSFP/vendor methods.
• The automated DCNN approach demonstrated applicability for LV chamber quantification and strain analysis in cardiotoxicity.

Conclusions:

• The developed DCNN tool enables automated and accurate LV chamber quantification and myocardial strain analysis from DENSE MRI.
• This automated approach is effective for monitoring cardiotoxicity in breast cancer patients undergoing chemotherapy.
• The findings support the integration of DCNN-based analysis for improved cardiac assessment in clinical oncology settings.