DeepColonLab: Attention Guided Separable Receptive Field Block Enhanced Deeplabv3+ Model for Colon Polyp Segmentation

Colon polyp segmentation is necessary for early colorectal cancer classification, helping to reduce deaths from one of the most common and deadly cancers worldwide. Accurate segmentation of colon polyp is difficult due to their diverse morphologies and different sizes. Existing models like convolutional neural networks may struggle to preserve fine-grained spatial details and transformer-based architectures may not be computational efficiency for real-time clinical use. So, we introduce DeepColonLab, a modification of the DeepLabV3+ model, specially designed for colon polyp segmentation. Our approach introduces a Separable Receptive Field Block (SRFB), inspired by human visual receptive fields, integrated with a Convolutional Block Attention Module (CBAM) to replace traditional Atrous Spatial Pyramid Pooling (ASPP). DeepLabV3+ is well-suited for colon polyp segmentation due to its encoder-decoder architecture and ASPP module, which enable effective multi-scale feature extraction and precise boundary delineation. It is lighter than many traditional segmentation models but can achieve high accuracy due to its structure. Original DeepLabV3+ with ASPP module lacks sufficient mechanisms for global context awareness and fine boundary refinement. Our proposed design enhances multiscale contextual information capture while preserving spatial resolution, particularly for small and irregularly shaped polyps. It enhanced receptive field's flexibility and better channel wise feature transformation for balancing efficiency and accuracy. Using lightweight EfficientNet encoders, DeepColonLab balances accuracy and computational efficiency. This model also provides better gradient flow and feature retention than base DeepLabV3+. This model outperformed most of the recent and state-of-the-art models on benchmark datasets—Kvasir, CVC-ClinicDB, and CVC-ColonDB— achieving Dice Coefficients of up to 0.9597 ± 0.0060 and Intersection over Union (IoU) scores of up to 0.9314 ± 0.0084. The efficiency of the model supports real-time medical imaging applications, making it a promising tool for clinical deployment in the management of colorectal cancer.