COLOR MEDICAL IMAGE VECTOR QUANTIZATION CODING USING K-MEANS & FUZZY C-MEANS: RETINAL AND CATARACT IMAGE

• Main Author: WAHYU SETIAWAN (NIM 23206023), AGUNG
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/7062
• Institution: Institut Teknologi Bandung

Retinal and cataract images play an important role in supporting medical diagnosis. Digital retinal and cataract images usually are represented in such a large data volume that efforts to access and display it takes a considerable amount of time. Digital medical image compression therefore become crucial in medical image transfer and its storage in electronic database server. This master thesis is concerned with the development of a color medical image coding scheme using vector quantization. K-means and fuzzy c-means are clustering techniques that are applied to create codebook in vector quantization (VQ) coding. This research investigates the performance of each of these clustering techniques in 5 color models: RGB, YUV 4:4:4. YUV 4:2:0, HSV and YIQ. The VQ coding scheme is conducted separately to image components in each channel of the color models. Reconstructed color image is obtained by combining the VQ decoding result of each image components. The VQ coding performance relies on the quality of the utilized code book. The VQ code book used in this research was developed from two kinds of training images, i.e. the retinal image and the cataract image. The retinal image was processed as a set of four quarter-images while the cataract image was processed as a set of two half-images. This special treatment is required to cope with the large computational load of the VQ code book generation, while ensuring the incorporation of the color and texture diversity of the training set in the resulted code book. Since fuzzy-c-means clustering technique is computationally more expensive than the k-means technique, a specific modification to the FCM is required. In this research, the fuzzy-c-means clustering is done separately to 16 equally-sized subimages to create 16 sub-code books. Training vectors from these 16 sub-code books are then unified as one single set of training data for a complete code-book generation. The RGB 444 color mode (coding of the red, green, and blue channels by the size of 4x4) produces the best subjective and objective quality of image coding. However, the optimum color models for teleophthalmology and electronic medical records are the YUV 4:2:0 and RGB 848 for retinal image, and YUV 4:2:0 for cataract image. Overall, the performance of k-means clustering technique is better than the FCM for color medical images, retinal, and cataract images. <br />