Optimize Multiscale Feature Hybrid-Net Deep Learning Approach Used for Automatic Pancreas Image Segmentation

The accurate and clean pancreatic segmentation problem is a significant and difficult challenge in the research ofmedical imaging. The proposed approach is helpful in Bottom-Up techniques for quickly and accurately identifyingpancreatic cancer-affected areas in the kidney and pancreas. The proposed approach is employed for multiple organdetection with fine outliers from Computed Tomography scan images with sharp precision. In the proposed approach,19 layers are used with 4 convolution layers. The optimized multiscale feature hybrid block is utilized. The convolutionlayer incorporates the small feature block with a similar or different size of kernel block. The dense convolution isdivided into sub-convolution layers with their own ”Batch normalization” layer. The convolution operation is performedon sub-convolution layers with the ”Swish activation” function. The aggregates function only uses features that arenecessary, causing unused data to be dropped at the convolution layer. The Validation Accuracy(98.82% ), precisionvalue (84.14±19.03) Training global accuracy (95.486%) values are improvised as compared to State-of-Art. Theproposed approach achieved a dice similarity index score upto 90.29±84.34%. During testing, the segmentation ofmedical images by the proposed approach, takes 1 to 2 seconds. In deep learning approaches, many convolution layersare added to the model to achieve high accuracy but training as well as unnecessary computational time also increases.The proposed approach overcomes this drawback with an optimized convolution layer number. The proposed approachis considering the only hybrid sub-convolution layer with required multiscale feature blocks only