Fully automatic lesion segmentation in breast MRI using mean-shift and graph-cuts on a region adjacency graph

Area of Science:

• Medical Imaging
• Radiology
• Computer-Aided Diagnosis

Background:

• Accurate segmentation of suspicious breast tissue in MRI is crucial for diagnosis.
• Existing methods often require manual input or lack comprehensive automation.
• Developing a fully automatic system can improve efficiency and consistency in breast MRI analysis.

Purpose of the Study:

• To present and evaluate a novel, fully automatic method for the detection and delineation of suspicious lesions in breast MRI.
• To assess the performance of this automated segmentation technique against expert-defined ground truth.

Main Methods:

• The method utilizes mean-shift clustering and graph-cuts on a region adjacency graph.
• Developed and tuned using multimodal (T1, T2, DCE-MRI) data from 35 subjects.
• Validated on two independent test sets (85 and 8 subjects) with manual 3D delineation as ground truth.

Main Results:

• Achieved 100% lesion detection rate with a mean of 4.5 ± 1.2 false positives per subject.
• Demonstrated a median Dice coefficient of 0.76 (Test set 1) and 0.75 (Test set 2).
• The false-positive rate is nearly 50% lower than previously reported fully automatic systems.

Conclusions:

• The proposed method is effective and accurate for automatic breast lesion segmentation in MRI.
• The system shows potential for cross-vendor application, enhancing its clinical utility.
• This represents a pioneering fully automatic approach for breast lesion detection and delineation in MRI.