Investigating the effect of Normalization Segmentation and Feature Selection techniques to improve classification performance of Deep Neural Net Models on Image datasets

Abstract

newlineThis work primarily focuses on developing classification and segmentation models for medical image datasets. The initial goal is to develop a predictive model for diagnostics of very prevalent three main kidney diseases: tumor, stone, and cyst, all of which are widespread and call for prompt medical attention. A comprehensive multi-class dataset containing 12,446 CT scan images of human kidneys, collected from various hospitals in Dhaka, Bangladesh using the Picture Archiving and Communication System (PACS), served as the basis for training and testing the classification model. Initial model development employed deep neural networks, specifically InceptionV3, ResNet50, and VGG16. Notably, using Adaptive Gradient Clipping (AGC) resulted in achieving top accuracies: 97.15% with VGG16, 99.5% with ResNet50, and 99.23% with InceptionV3. Overall, ResNet50 and InceptionV3 demonstrated superior classification performance when combined with the Adaptive Gradient Clipping strategy. newlineThe second work is to design a segmentation and classification model for a Diabetic Retinopathy (DR) image dataset. DR is a severe condition affecting diabetic individuals, resulting from hemorrhages in the light-sensitive retinal region. Accurate diagnosis is challenging due to a number of issues with current AI-driven DR diagnosis techniques, such as noise, uneven lighting, and low contrast. To enhance filtering performance, we developed an Adaptive Gabor Filter (AGF) which is a linear filter and used the Chaotic Map. Features were extracted using Local Binary Patterns (LBP) which is based on texture features, Speeded-Up Robust Features (SURF) which identifies key points using feature vector, and Texture Energy Measurement (TEM). A dense block from DenseNet, an Attention layer, and an Optimized Gated Recurrent Unit (OGRU) supplemented by a Self-Adaptive Northern Goshawk Optimization (SANGO) method are all used in this phase to boost classification performance. newlineThree datasets were used to evaluate the system: DiaRetDB1, APTOS 2019, and EyePacs. The intention of using three different datasets is to determine the robustness of the proposed model. Experimental Results showed that the system was reliable and resilient. Additionally, the Grad Cam in the recommended method ensures that segmentation and classification performance are implemented well. Accuracy, Precision, Recall, F1-Measure, Intersection over Union (IOU) which is a measure of overlap of predicted image and ground truth, and Dice Similarity Coefficient (DSC) which measures similarity between two sets of images, and other measures are used to illustrate performance. Additionally, the outcome performance is analyzed using the five-fold categorization. The proposed model s accuracy on the DiaRetDB1 dataset was 99.01%, on the APTOS 2019 dataset it was 98.98%, and on the EyePacs dataset it was 99.12%. newline newline