Evaluation of the Effect of Steganography on Medical Image Classification Accuracy
Predictors and features that are used in teleradiology and machinebased auto diagnosis in medicine are often not put into consideration while evaluating medical image Steganography algorithms. In this paper, the effect of embedded security data in automated diagnosis was evaluated using Support Vector Machine (SVM) image classification of Chest X-rays Scan of Normal and Pneumonia patients. The goal is to quantify and qualify disease classification parameters because of the addition of steganographic security data in to the image. Four textural image features: Contrast,Homogeneity Energy, and Entropy were used as medical image biomarkers. Their statistical properties for the disease conditions (normal or pneumonia)were profiled and used in SVM training. The evaluation parameters for the machine learning models include accuracy, specificity, recall, and precision. The baseline (before the addition of security data)performance was 86.14% accuracy,82.18% recall, 90.10% specificity, and 89.25% precision while a typical performance after the addition of security data was 82.18% accuracy, 85.15% recall, 79.21% specificity, and 80.37% precision.We conclude that embedding strength, watermark payload, and region of embedding should be carefully selected to avoid the automated diagnostic outcome by a steganographic security changing algorithm algorithm to a medical image.