ROI-AWARE DEEP LEARNING FRAMEWORK FOR BRAIN MRI COMPRESSION

Abstract

Magnetic Resonance Imaging (MRI) has become a routine tool for diagnosing brain conditions, especially tumours. One thing that comes with MRI's clinical usefulness is a real practical problem: the scans take up a lot of storage. A single patient study can run into hundreds of megabytes, and hospitals deal with thousands of such studies every year. What makes medical compression different from regular image compression is that we cannot afford to blur out the very region a radiologist is going to look at. So we need a compression scheme that knows what to protect and what to compress. In this thesis, we begin by surveying the existing landscape of image compression, both classical (JPEG, JPEG2000) and learning-based (autoencoder-style codecs), and the work that has been done on Region-of-Interest (ROI) compression for medical images. Following this, we present a deep-learning-based ROI-aware compression pipeline for brain MRI. The pipeline has three connected stages: a U-Net segmentation model identifies the tumour region from each axial FLAIR slice, the predicted mask is expanded by a controllable safety margin to absorb small segmentation errors, and a convolutional autoencoder compresses everything outside the safe region while the ROI itself is kept intact. The final image is reconstructed by stitching the original ROI back with the decoded background. Next, we add two ideas that make the framework more practical. We define an adaptive compression-ratio rule that scales the background CR (16:1, 32:1 or 64:1) based on the relative ROI size in each slice. And we run a margin sweep over m = 0, 2, 4, 6, 8 and 12 pixels to map how the trade-off between compression efficiency and diagnostic safety actually behaves as the ROI is expanded. The second of these is also the central contribution of the conference paper communicated from this work, for which we revised the autoencoder design and trained it with an ROI-aware loss. Finally, we report what the framework actually achieves on the BraTS 2020 dataset. The U-Net reaches a mean Dice of 0.7795 and recall of 0.8440. The full system achieves a PSNR of 29.78 dB, SSIM of 0.8753, a mean effective compression ratio of 25.50:1 with a maximum of 64:1, all while keeping the tumour region bit-exact by construction. The margin sweep shows that ROI coverage improves from 0.756 (m = 0) to 0.982 (m = 12), with leakage dropping from 0.244 to 0.018 at a controlled cost to global PSNR. The practical operating range works out to m = 6 to 8 pixels. We close the thesis with a discussion of limitations, concrete future directions, and the social implications of deploying this kind of system in real clinical settings.