Breast cancer detection in mammogram: combining modified CNN and texture feature based approach

Customized deep neural networks are being used to assess medical imaging and pathology data. The proper assessment of malignancy using digital mammography images is a challenging task. This study implements a system for autonomously diagnosing cancer using an integration method, which includes CNN and image texture attribute extraction. The nine-layer customized convolutional neural network is used to categorize data in the CNN stage. To improve the effectiveness of categorization in the extraction-based phase, texture features are defined and their dimension is reduced using Uniform Manifold Approximation and Projection (UMAP). The findings of each phase were combined by an ensemble algorithm to arrive at the ultimate conclusion. The final categorization is presumed to be malignant if any of the stage's output is malignant. On the MIAS repository, our ensemble method's testing specificity and accuracy are 97.8% and 98%, respectively, while on the DDSM repository, they are 98.3% and 97.9%. The combination method improves measurement metrics across each phase independently, as per the experimental findings.

Automated summary

Customized deep neural networks and image texture attribute extraction are used in this study to autonomously diagnose cancer based on digital mammography images. The findings show that the combination method improves the accuracy and specificity of classification.