Deep Augmented Metric Learning Network for Prostate Cancer Classification in Ultrasound Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Prostate cancer screening commonly uses expensive MRIs and invasive biopsies.
• Transrectal ultrasound offers a cost-effective, non-invasive alternative but struggles with distinguishing benign from malignant tumors due to image similarities.

Purpose of the Study:

• To introduce a novel Deep Augmented Metric Learning (DAML) network for improved prostate cancer classification using ultrasound imaging.
• To enhance the differentiation of subtle features in ultrasound images for more accurate auxiliary diagnosis.

Main Methods:

• Developed the Deep Augmented Metric Learning (DAML) network with Semantic Differences Mining Strategy (SDMS) for feature differentiation.
• Integrated Linear Interpolation Augmentation Strategy (LIAS) and Permutation-Aided Reconstruction Loss (PARL) to address class variability and limited samples.
• Conducted comprehensive empirical assessments comparing DAML against other models.

Main Results:

• The DAML network achieved high classification accuracies: 0.857 for benign and 0.888 for malignant prostate cancers.
• The proposed strategies (SDMS, LIAS, PARL) effectively enhanced feature representation and addressed class variability.
• Empirical assessments validated the superior performance of the DAML model.

Conclusions:

• The DAML network demonstrates significant potential for accurate and efficient prostate cancer classification using ultrasound imaging.
• This approach offers a promising advancement in non-invasive diagnostic tools for prostate cancer, overcoming limitations of current methods.