A Data-Efficient Machine Learning Approach for Breast Ultrasound Lesion Classification Integrating Image-Derived...

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Breast ultrasound is crucial for lesion evaluation, but heterogeneity and small datasets hinder deep learning.
• Current methods struggle with reliable lesion characterization and generalizability in routine practice.

Purpose of the Study:

• To evaluate a data-efficient framework integrating image features and clinical descriptors for improved breast ultrasound lesion classification.
• To enhance diagnostic performance in small cohort settings.

Main Methods:

• Extracted image features using a pretrained convolutional neural network (CNN).
• Combined image features with manual sonographic descriptors into a unified tabular dataset.
• Trained gradient-boosted tree models with descriptor-only and fused feature sets, validated using cross-validation and an external test set.

Main Results:

• Descriptor-only models achieved 0.88 accuracy and 0.95 AUC.
• The fused framework reached 0.88 accuracy and 0.96 AUC, with 1.00 sensitivity for malignant lesions.
• The fused approach showed more stable generalization, particularly for malignant cases.

Conclusions:

• Integrating image-derived features with clinical descriptors in a tabular learning framework offers a robust, data-efficient method for breast ultrasound lesion classification.
• This strategy aids decision-making in small datasets and is practical when large deep learning models are infeasible.