A statistical method for evaluation quality of medical images: a case study in bit discarding and image compression

Abstract

Many studies have been performed on quality evaluation for subtle differences in medical images. However, only limited success has been achieved. In this paper, medical images were prior manipulated by denoising, lossy compression and filtering. The Moran statistics is then applied to extract spatial information of images and using Kolmogorov–Smirnov (KS) test to determine whether the manipulated and original images differ significantly. Results show that on average discarding 1–2 bits in T1 and CR images or 2–3 bits in T2 and body CT images are indistinguishable. This method is also applied to a reconstructed MR, body CT image and an electronic SMPTE (Society of Motion Picture and Television Engineer) phantom from lossy image compression software. Compression ratios of 16:1 for a MR image, 8–9:1 for a cropped body CT image, 7:1 and 5:1 for high- and low-resolution regions in electronic phantom is proved undifferentiated from original. The proposed method is useful for complementing the human visual system, to optimize the performance of image compression technique.

Introduction

Medical images are often deteriorated by noise due to various interference sources and other phenomena that affect the measurement process in imaging acquisition systems. If the small difference between normal and abnormal tissues is contaminated by noise, direct analysis of the acquired images is often more difficult [1]. Noise in medical images also degrades the correlation between pixels. As a result, the reversible image compression ratio can only achieve a maximum compression ratio of about 2–4 [2].

Steward et al. [3] suggested that 5–6 least significant bits (LSBs) are spectrally white in an image by applying the power spectrum method to verify each bit plane in a magnetic resonance (MR) image. They concluded that the gray level dynamic range of a MR image spanned only six or seven bits at most. Chan et al. [4] showed that the detection accuracy was independent of pixel depth for discarded up to three LSB planes from a digitized mammography with 12 bits depth.

Olsen [5] estimated the standard deviation (STD) for Gaussian and uniform distributed white additive noise in images. He found that the use of STD to estimate noise in the residual images was most reliable. The residual image was obtained by subtracting the pre-filtered image from the original image, where smoothed image is obtained by average or median filters. Chuang et al. [6], [7], [8] reported that 8–9 information bits exist in medical images and 4–5 bits contain only noise. Discarding these noisy bits would not visually deteriorate image quality, as indicated in their study. However, the true effect on the image has not been verified.

With the evolution of teleradiology, the use of lossy compression technique is a requirement for reducing the image volume and speeding up communication [9]. However, greater compression is obtained at the expense of image quality. Efforts have been devoted to evaluate the subtle differences between images using visual or numerical methods. Many studies had tried to build a visual threshold or acceptance level of lossy image compression. However, the radiological tasks is widely varied and the building of visual threshold is task dependent and time consuming [2], [10], [11], [12], [13]. The normalized root mean square error (NMSE) or mean square error (MSE) for numerical estimation was most commonly used to measure the degradation in reconstructed images. The guidance document for Picture Archiving and Communication System (PACS) also requires that manufacturers report to FDA, the NMSE of their lossy compression techniques [14]. Nevertheless, NMSE and MSE are found to neither correlate nor provide any quality information [2], [10], [13], [15], [16].

The objective image quality metrics, visual discrimination model (VDM) and visible differences predictor (VDP), has been developed in past years. These metrics combined perceptual quality measurement by taking into account human visual system (HVS) features [16], [17], [18], [19], [20], [21], [22]. Since a human observer is the end user of image quality measurement, image quality model based on HVS seemed to be more appropriate for user perceptual. Recently, the detection threshold of artifact in medical image subject to JPEG and JPEG 2000 compression was reported [23]. They mentioned that the VDM predictions of artifact visibility were highly correlated with observer performance. However, it was reported [21], [22], [24] that none of the objective quality model outperformed others by comparing nine proponent objective quality models and all of them were statistically indistinguishable with the peak signal to noise ratio (PSNR).

Recently, Wang and Bovic [25] suggested a Q index to estimate image quality. The image spatial information is estimated from a local region with a 8×8 window size instead of from pixel, as indicated by their study. They showed that the variation of the Q index for a manipulated image was accordant with subjective image quality metrics. Chen et al. [26], [27] applied the Moran statistics in a local region to extract image spatial information. The sharpness or smoothness in image is indicated well by this statistics [27]. In this study we propose to evaluate spatial information in image by using Moran statistics and then a statistical method is applied to determine if this subtle difference between images is due to chance.