A comparative study of image compression (low vs high) on fuzzy logic and ANN based medical DEXA image's classification

The objective work of this paper is to gather DEXA medical images, make a comparative study of their classification to Normal and clearly abnormal (Osteoporosis) images, due to low and high compression, using image features, Fuzzy Logic and Artificial Neural Networks (ANN). Compression is applied using quantization (Q), which determines the Compression Ratio (CR), Root Mean Squared Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR). With Q < 1, a low compression ratio of 56%, RMSE of 5.8% and PSNR of 32.9 are achieved, while Q > 10 results in a high compression ratio of 95%, RMSE of 15.2% and a decreased PSNR of 24.4. The results show that low compression causes feature values to converge, while high compression leads to divergence, which is considered a core contribution of this work. Compression reduced the number of rules in the Fuzzy Inference System (FIS) in spite of successfully classifying most images. ANN training became faster and more efficient, while testing performance decreased compared to uncompressed images due to feature convergence caused by compression. Overall, compression reduced classification accuracy from 94.28% in uncompressed images, to 93.47% with high compression, and 89.47% with low compression.