[Paper Review] Neural networks approach for mammography diagnosis using wavelets features

Essam A. Rashed, Mohamed M. Awad|arXiv (Cornell University)|Mar 6, 2020

AI in cancer detection18 references4 citations

TL;DR

This paper proposes a supervised mammography diagnosis system using wavelet-based feature extraction and artificial neural networks. Multilevel wavelet decomposition transforms digital mammograms into feature vectors, which are classified by a neural network to distinguish tumor types and risk levels, achieving promising results with radiologist-labeled data, particularly through enhanced feature representation via multilevel decomposition compared to single-level methods.

ABSTRACT

A supervised diagnosis system for digital mammogram is developed. The diagnosis processes are done by transforming the data of the images into a feature vector using wavelets multilevel decomposition. This vector is used as the feature tailored toward separating different mammogram classes. The suggested model consists of artificial neural networks designed for classifying mammograms according to tumor type and risk level. Results are enhanced from our previous study by extracting feature vectors using multilevel decompositions instead of one level of decomposition. Radiologist-labeled images were used to evaluate the diagnosis system. Results are very promising and show possible guide for future work.

Motivation & Objective

• To develop a supervised diagnosis system for digital mammograms using advanced feature extraction.
• To improve classification performance by replacing single-level with multilele wavelet decomposition for feature extraction.
• To leverage artificial neural networks for accurate classification of tumor types and risk levels in mammograms.
• To validate the system using radiologist-annotated mammographic images for clinical relevance.
• To provide a foundation for future research in automated breast cancer diagnosis using wavelet and neural network integration.

Proposed method

• Multilevel wavelet decomposition is applied to digital mammogram images to extract detailed spatial-frequency features.
• The resulting wavelet coefficients are combined into a feature vector representing each image.
• The feature vector is fed into a trained artificial neural network for classification.
• The neural network is trained to classify mammograms into distinct tumor types and risk levels.
• The system uses radiologist-labeled data for supervised learning and performance evaluation.
• The approach improves upon prior work by using multilevel decomposition instead of single-level decomposition for richer feature representation.

Experimental results

Research questions

• RQ1 Can multilevel wavelet decomposition improve feature representation for mammogram classification compared to single-level decomposition?
• RQ2 How effective is a neural network in classifying mammograms into tumor types and risk levels using wavelet-derived features?
• RQ3 To what extent does the proposed system achieve high diagnostic accuracy using radiologist-labeled data?
• RQ4 What is the potential of combining wavelet transforms with neural networks for automated mammography diagnosis?
• RQ5 How does the system's performance compare to previous approaches using simpler feature extraction methods?

Key findings

• The use of multilevel wavelet decomposition significantly enhances feature representation compared to single-level decomposition, improving classification performance.
• The proposed neural network model achieves promising classification accuracy in distinguishing tumor types and risk levels in mammograms.
• Results demonstrate the effectiveness of wavelet-based features in capturing relevant diagnostic patterns from digital mammograms.
• The system shows strong potential as a decision-support tool in clinical settings, particularly with radiologist-validated data.
• The study provides a foundation for future work in automated breast cancer diagnosis using hybrid wavelet-neural network approaches.
• The model outperforms previous versions of the system due to improved feature extraction through multilevel decomposition.