OVGGNet: Optimized deep learning for lesion segmentation of medical images using color features

Abstract

Medical image segmentation is a challenging task in image processing, automatic segmentation needs expert suggestions and clinical practices such as treatment planning, disease diagnosis and disease progression. The primary problems in inaccurate lesion segmentation are variation of lesion conditions, lesion irregularity and presence of similarity between lesion regions. To avoid these problems, the Optimized Deep Learning model named as OVGGNet is proposed for the accurate lesion segmentation of medical images. Different data sources are used to collect different kinds of medical images that are preprocessed to enhance contrast of the medical images. The contrast enhanced images are predicted from this preprocessed stage, and then the Dilated dense VGG19 (DDVGG19) model is employed to extract global and local features from the enhanced images. In the feature extraction stage, the Shi Tomasi Corner detector is applied for shape identification and the linear dimensionality reduction approach is applied for undesired information removal. The Optimized Color Feature (OCF) operation is included at the final lesion segmentation tasks to carry out the different kinds of segmented output. Finally, the mapped images are predicted by fusing original and segmented images and its visual representations are provided in the experimental validation. The numerical and graphical validations are conducted to find the optimal performances during lesion segmentation of medical images. The significant performance measures such as segmentation accuracy and Dice Similarity Coefficient index (DSC) achieved performances of 98.81 % and 96.4 % respectively. The comparative analysis provided that the proposed model attained a better performance rather than existing approaches.