Deep transfer learning based hierarchical CAD system designs for SFM images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate classification of mammographic masses is crucial for effective breast cancer diagnosis.
• Hierarchical classification frameworks offer a structured approach to analyzing complex medical data.
• Deep transfer learning models show promise in medical image analysis tasks.

Purpose of the Study:

• To develop and evaluate a hierarchical deep transfer learning framework for classifying mammographic masses.
• To compare the performance of different deep transfer learning models within a hierarchical structure.
• To create a hybrid hierarchical computer-aided diagnosis (CAD) system for improved mass classification.

Main Methods:

• Utilized the DDSM dataset comprising 518 mammographic mass images across benign, suspicious, and malignant categories.
• Employed a two-node hierarchical classification framework: Node 1 classified masses as probably benign or suspicious abnormality; Node 2 further classified suspicious masses.
• Experimented with deep transfer learning models including VGG16, VGG19, GoogleNet, and ResNet50 for feature extraction and classification.

Main Results:

• The VGG19 model achieved 93% accuracy at Node 1, and VGG16 achieved 90% accuracy at Node 2.
• A hybrid hierarchical CAD system combining VGG19 (Node 1) and VGG16 (Node 2) yielded an 88% overall accuracy.
• A further refined hybrid system using VGG19/ANFC-LH and VGG16/ANFC-LH classifiers achieved a highest classification accuracy of 92%.

Conclusions:

• The proposed hybrid hierarchical CAD systems demonstrate significant potential for accurate, step-wise classification of mammographic masses.
• Deep transfer learning models, when integrated into a hierarchical framework, can enhance diagnostic capabilities in mammography.
• The findings suggest the clinical utility of these advanced AI systems for improving breast cancer detection and diagnosis.