Hybrid Region of Interest Based Near-Lossless Codec for Brain Tumour Images Using Convolutional Autoencoder

One of the most significant industries producing digital images worldwide is radiology. The advancements in radiological equipment have ensured the production of high-definition digital medical images. Irrespective of the growth in storage and network facilities, these high-definition images still face the problem of high storage space requirements and high transmission costs. To handle aforementioned problems and pave way for faster, hassle-free transmission and effective storage of such medical images for telemedicine services, we need enhanced image compression techniques. Medical images tend to have data with both high and low clinical importance for downstream analysis and treatment. Compression algorithms must be developed in order to handle both high and low clinically important data at the same time to improve the compression standard of medical images. We propose a Convolutional autoencoder technique for Region of Interest based hybrid near-lossless medical image compression aided by “You Only Look Once” (YOLO) deep learning algorithm. This work aims to achieve ROI-based near-lossless compression with notable compression ratio and medical image quality. To achieve this ROI-based near-lossless compression, we employed a combination of YOLO object detection algorithm, Convolutional autoencoder, and Haar wavelet transform with SPIHT encoding on grayscale Magnetic Resonance brain tumour images. The proposed approach was evaluated against several existing standard compression methods. Results inferred that our proposed method assured the near-lossless image compression scenario by maintaining the quality of medical images after decompression and comparatively reduced the storage and transmission cost by ensuring an effective compression ratio.