ESE and Transfer Learning for Breast Tumor Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer molecular subtype recognition is crucial for treatment.
• Accurate and non-invasive diagnostic tools are needed.
• Deep learning shows promise in medical image analysis.

Purpose of the Study:

• To propose TLese-ResNet, a lightweight neural network for breast cancer molecular subtype recognition.
• To evaluate the model's performance on mammographic images.
• To provide an effective auxiliary tool for clinicians.

Main Methods:

• Developed TLese-ResNet using inverted residual network, efficient squeeze excitation (ESE) module, and double transfer learning.
• Utilized a dataset of mammography images (CC and MLO views).
• Employed data augmentation and a two-stage transfer learning process (ImageNet -> COVID-19 X-ray -> mammography).

Main Results:

• TLese-ResNet achieved a mean accuracy of 0.818 and an area under the curve of 0.883 on mammographic images.
• The model outperformed state-of-the-art deep learning models like ResNet-50 and DenseNet-121.
• Five-fold cross-validation was used to assess performance.

Conclusions:

• TLese-ResNet is an effective and non-invasive auxiliary tool for breast cancer molecular subtype identification.
• The proposed architecture offers reduced complexity and enhanced feature expression.
• This model can assist clinicians in making informed diagnostic decisions.