Apply ML techniques leveraging spatial and frequency features for comprehensive medical image analysis

Abstract

Accurate analysis of medical images is crucial for early diagnosis and treatment planning in healthcare. In the past it has been the case that we mainly see two approaches spatial which looks at pixel intensity and texture and frequency which we get from transforms like the Discrete Wavelet Transform out in the frequency domain. But the issue with that is we are often limited in what a model is able to do diagnosis wise because we aren’t representing the full picture. In this work we put forth a full machine learning based solution which brings together spatial and frequency features for better medical image analysis. We use a custom made Convolutional Neural Network for the extraction of spatial features which in turn present local and structural information. At the same  time, we use DWT to obtain frequency features which we use for high frequency elements and textural variation
across many scales. We then put these two feature sets together and run them through Principal Component Analysis for dimensionality reduction. We use this hybrid feature set to train many classifiers which include Support Vector Machine, Random Forest and a CNN-MLP hybrid. We evaluated our model on standard sets of images from Brain MRI and Chest X Ray. What we found is that our combined model does better in terms of accuracy and sensitivity then models which use only one domain. We also see that the put together use of spatial and frequency features improve diagnostic performance which we think has great promise for use in clinical diagnostic tools.