Python code for deep residual network with multi-level feature fusion and genetic optimization for endoscopic polyp segmentation

Abstract：Early screening for colorectal cancer heavily relies on colonoscopy. However, traditional diagnostic paradigms suffer from high polyp miss rates and dependence on physicians' subjective experience. Deep learning-based analysis of endoscopic images offers an effective means for computer-aided diagnosis. Deep residual networks, in particular, have attracted significant attention due to their powerful feature extraction capabilities. Nevertheless, model performance is often constrained by the tedious and suboptimal process of manual hyperparameter tuning. To address these challenges, this paper proposes an automatic polyp detection and segmentation method for endoscopic images that integrates deep residual networks with a genetic algorithm (GA). Firstly, to tackle the issue of spatial information loss caused by the simple decoder structure in baseline segmentation models, an improved U-Net-like architecture is constructed. This architecture introduces skip connections with dynamic channel adaptation and size alignment mechanisms, builds a four-level feature pyramid, and adopts a progressive upsampling strategy. This enables effective fusion of fine-grained details from the encoder with semantic information from the decoder, facilitating multi-scale feature reconstruction. Building upon this, an automated hyperparameter optimization scheme based on a genetic algorithm is designed and implemented. This scheme employs a hybrid encoding strategy and uses the Intersection over Union (IoU) on the validation set as the fitness function to perform a global search in the parameter space for the optimal hyperparameter combination. Experimental results on the Kvasir-SEG dataset demonstrate that the improved model significantly outperforms the baseline segmentation model, increasing IoU from 0.4578 to 0.9311. After optimization via the genetic algorithm, the model's performance is further enhanced, achieving an IoU of 0.9541. Furthermore, the optimized model produces smoother segmentation boundaries and effectively reduces false positives and missed detections. This method provides a novel technical pathway for intelligent and precise analysis of endoscopic images, demonstrating significant potential for applications in computer-aided diagnosis.