Dual-Intended Deep Learning Model for Breast Cancer Diagnosis in Ultrasound Imaging.

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Medicine
• Oncology and Radiology

Background:

• Automated medical data analysis is crucial for reliable cancer diagnosis and prognosis.
• Breast cancer screening in ultrasound imaging requires accurate and efficient methods.
• Radiomic feature extraction aids in characterizing lesions but can be high-dimensional.

Purpose of the Study:

• To propose an automated breast cancer screening method using deep learning on ultrasound images.
• To develop a model for simultaneous segmentation of breast lesions and radiomic feature extraction.
• To evaluate the diagnostic performance of extracted radiomic features for malignant lesions.

Main Methods:

• A convolutional deep autoencoder model was employed for lesion segmentation and deep-radiomic extraction (4 features).
• High-dimensional conventional imaging features (354) were reduced to 12 radiomics using spectral embedding.
• A random forest model was trained and validated on 780 ultrasound images (benign, malignant, normal) for binary classification.

Main Results:

• The deep autoencoder performed simultaneous segmentation and low-dimensional deep-radiomic extraction.
• Dimensionality reduction techniques were applied to conventional imaging features.
• The maximal cross-validated random forest model achieved a binary classification accuracy of 78.5% (65.1-84.1%) for diagnosing malignant lesions.

Conclusions:

• The proposed automated method effectively segments breast lesions and extracts relevant radiomic features from ultrasound images.
• Combining deep-radiomic and spectrally embedded features shows promise for improving breast cancer diagnosis.
• The developed deep learning approach offers a potential tool for automated breast cancer screening.