Improved diagnostic model for gastrointestinal disorders through novel masking and semantic segmentation

Abstract

The integration of advanced computing techniques, such as computational intelligence, in disease classification and prediction through medical imaging has become a prominent practice within the field of computer vision. A critical aspect of this process is precise image segmentation, which helps identify and isolate relevant areas in medical images, thereby streamlining the analysis and enhancing the accuracy of predictions. Among various segmentation methods, semantic segmentation, which assigns class labels to each pixel, has proven particularly effective. In this study, we introduce the U+EfficientB0 Net model, specifically designed to segment gastro-intestinal images, including dyed-lifted polyps, polyps, and ulcerative colitis. The proposed work is different from traditional EffeicientNetB0 encoder as it utilizes Decoder-guided Adaptive Skip Refinement (DASR) module for feature refinement before combining the features obtained from encoder with the features obtained through decoder. Utilizing advanced computing techniques like computational intelligence for disease classification and prediction through medical imaging is a prominent practice in computer vision. Precise image segmentation is crucial as it identifies relevant areas, streamlining analysis, and bolstering prediction accuracy. Semantic segmentation, assigning class labels to each pixel, proves especially effective. This study introduces the (U+EfficientB0) Net model specialized in segmenting gastro-intestinal images, notably dyed-lifted-polyps, polyps and ulcerative colitis leveraging the Kvasir-Seg dataset for polyps and while corresponding masks for dyed-lifted polyps and ulcerative colitis are generated using pixel-based annotation approach. Trained on a thousand images supervised by medical experts, proposed model achieved remarkable accuracy (92.89%) and dice coefficient (80%). Additionally, the study presents a classification approach using DenseNet-121 that uses the predicted segmented images as input and its performance is compared with similar model’s using non-segmented image and achieved better results (Accuracy 99.63) over segmented images in earlier epochs.