Wavelet leader based formalism to compute multifractal features for classifying lung nodules in X-ray images

TL;DR

This paper introduces a novel lung nodule classification method using wavelet leader multifractal features combined with classical textures, achieving improved accuracy and efficiency over existing wavelet-based approaches.

Contribution

It presents a new multifractal feature extraction formalism based on wavelet leaders for lung nodule classification in X-ray images.

Findings

Maximum ROC AUC of 75% with combined features

Improved accuracy and efficiency over previous wavelet methods

Data augmentation and parameter tuning enhance results

Abstract

This paper presents and validates a novel lung nodule classification algorithm that uses multifractal features found in X-ray images. The proposed method includes a pre-processing step where two enhancement techniques are applied: histogram equalization and a combination of wavelet decomposition and morphological operations. As a novelty, multifractal features using wavelet leader based formalism are used with Support Vector Machine classifier; other classical texture features were also included. Best results were obtained when using multifractal features in combination with classical texture features, with a maximum ROC AUC of 75\%. The results show improvements when using data augmentation technique, and parameter optimization. The proposed method proved to be more efficient and accurate than Modulus Maxima Wavelet Formalism in both computational cost and accuracy when compared in a similar experimental set up.