A Comparison of Lossless Image Compression Algorithms for Colour Retina Images

Diabetic retinopathy is the leading cause of blindness in the adult population. In order to effectively identify patients suffering from the disease, mass-screening efforts are underway during which digital images of the retina are captured and then assessed by an ophthalmologist. In order to identify features such as exudates and microaneurysms, which are typically very small in extent, retinal images are captured at high resolutions. This in turn means large file sizes and, considering the archival of typically thousands of records, a high demand on computational resources, in particular storage space as well as bandwidth when used in a Picture Archiving and Communications System (PACS). Image compression therefore seems a necessary step. Image compression algorithms can be divided into two groups: lossy techniques where some of the visually less important image data is discarded in order to improve compression ratios, and lossless methods which allow the restoration of the original data. As the features that indicate retinopathy are very small in size and following legislation in several countries, only lossless compression seems suitable for retinal images. In this article, we present experiments aimed to identify a suitable compression algorithm for colour retina images (Schaefer & Starosolski, 2006). Such an algorithm, in order to prove useful in a real-life PACS, should not only reduce the file size of the images significantly but also has to be fast enough, both for compression and decompression. Furthermore, it should be covered by international standards such as ISO standards and, in particular for medical imaging, the Digital Imaging and Communication in Medicine (DICOM) standard (Mildenberger, Eichelberg, & Martin, 2002; National Electrical Manufacturers Association, 2004). For our study, we therefore selected those compression algorithms that are supported in DICOM, namely TIFF PackBits (Adobe Systems Inc., 1995), Lossless JPEG (Langdon, Gulati, & Seiler, 1992), JPEG-LS (ISO/IEC, 1999), and JPEG2000 (ISO/IEC, 2002). For comparison, we also included CALIC (Wu, 1997), which is often employed for benchmarking compression algorithms. All algorithms were evaluated in terms of compression ratio which describes the reduction of file size and speed. For speed, we consider both the time it takes to encode an image (compression speed) and to decode (decompression speed), as both are relevant within a PACS. Experiments were performed on a large dataset of more than 800 colour retinal images, which were also divided into subgroups according to retinal region (nasal, posterior, and temporal) and images size. Overall, JPEG-LS was found to be the best performing algorithm as it provides good compression ratios coupled with high speed.