Key Technology Research Of Medical Image-assisted Diagnosis Based On Texture Analysis

With the development of medical imaging technology,the number of medical images has increased on a large scale.These medical images assist radiologists in disease diagnosis but also increase the radiologists’ workload.In traditional medical image diagnosis,radiologists mainly observe a set of two-dimensional slice images to find lesions.It often depends on the experience of the radiologist.As for accurate discoveries of mechanism changes inside lesions and relationships with surrounding biological tissues,etc.,they are difficult to achieve.Texture as a mathematical descriptive property,can easily provide quantitative measurements of lesion characteristics.Medical image-assisted diagnosis based on texture analysis is a computer-aided diagnosis system for quantitative data analysis.It is the main research content of radiomics.In recent years,radiomic research based on texture features has become the main direction of medical image-assisted diagnosis.Traditional texture analysis methods are usually difficult to quantify regions of interest in medical images due to low spatial resolutions,complex internal structures of lesions,volumetric effects and ghosting interferences,etc.This paper systematically analyzed the crucial techniques of texture enhancement,texture feature extraction,and texture classification involved in medical image-assisted diagnosis based on texture analysis,summarized the advantages and disadvantages of the existing related technologies,and proposed some solutions in combination with characteristics of medical images.These solutions have achieved good results,including:1.A wavelet transform-based fractional differential algorithm applied to texture enhancement of regions of interest in medical images was proposed.The main purpose of texture enhancement is to highlight high-frequency contour information with more grayscale changes and quicker changes,and to preserve low-frequency smoothing information.The algorithm used wavelet transform to separate high-frequency and low-frequency components of a region of interest,and constructed a mask based on the Grumwald-Letnikov definition to convolve the high-frequency components.The mask was a multi-directional symmetrically enhanced fractional difference operator with a compensation parameter.Perfect reconstruction characteristics of wavelet inverse transform enabled the modification of transform coefficients(i.e.components)to be remarkable in a reconstructed image.The experimental results showed that the algorithm preserved low-frequency smooth texture information while enhanced high-frequency contour information,it made the texture information of a region of interest richer and the internal details clearer.The classification experiments of regions of interest before and after enhancement showed that the algorithm was also useful for the assisted diagnosis based on texture feature classification.2.A theoretical framework of multi-layered texture analysis combining multi-resolution analysis and statistical analysis was proposed,and three texture feature extraction algorithms for regions of interest of medical images were designed and implemented.The main purpose of texture feature extraction is to find stable descriptors with similar properties between same type of samples,and these descriptors should have significant differences between different types of samples.First,this theoretical framework used a multi-resolution analysis method to capture high-frequency detail components of a region of interest(characteristics of lesions that are difficult to visually distinguish in medical images are likely to be embedded in high-frequency detail components;a detail component is a coefficient matrix).Second,in training samples,as for a detail component,it performed rule analysis on the coefficient matrices that corresponding to this detail component.Then,as for a new sample,it discretized the coefficient matrix that corresponding to this detail component,and extracted mathematical descriptors of the discretized coefficient matrix using statistical methods.This paper designed and implemented three texture feature extraction methods:(1)A multi-resolution statistical analysis method based on linear normalization and flooring.This algorithm divided element values of a coefficient matrix into N equal-width intervals to discretize the coefficients.(2)A multi-resolution statistical analysis method based on minimum and maximum.In training samples,with respect to a detail component of the multi-resolution analysis,which is represented in a particular scale and a particular direction,this algorithm calculated a minimum and a maximum from the coefficient matrices that corresponding to the detail component,and discretized the coefficients based on N equal-width intervals between the minimum and the maximum.(3)A multi-resolution statistical analysis method based on the mean minimum and mean maximum.In training samples,with respect to a detail component of the multi-resolution analysis,which is represented in a particular scale and a particular direction,this algorithm calculated an averaged minimum and an averaged maximum from the coefficient matrices that corresponding to the detail component,this algorithm translated the coefficient matrix that corresponding to the detail component based on the averaged minimum,then divided N equal-width intervals based on the averaged minimum and the averaged maximum to discretize the coefficients,and constrained the coefficients that falling outside the N equal-width intervals.The above three algorithms all used statistical methods to extract mathematical descriptors(as texture features)of the discretized coefficient matrices.The experimental results showed that the extracted texture features had better performance than traditional texture analysis methods in medical image-assisted diagnosis based on texture analysis.These texture features could potentially enhance radiologists’ visual diagnostic capabilities.This may be related to diagnostic problems but not necessarily visually visible.3.A multi-group texture feature composite classification model based on category statistics was proposed.The model calculated category attribute probabilities of a new sample based on statistical characteristics of each component in a multi-resolution analysis.First,the model grouped training samples by category.In each group,it performed statistics on coefficients of components.Then,with respect to a coefficient matrix corresponding to a new sample,discretization was performed based on each group of statistical results,and texture features were extracted on the discretized coefficient matrices to obtain multiple groups of texture features.As for multiple groups of texture features for a new sample,this model calculated the probabilities that each group of features belongs to each category,and finally made a decision.The experimental results showed that this model improved the classification performance compared to traditional texture classification models.