2MGAS-Net: A multi-level multi-scale gated attentional squeezed network for polyp segmentation Ibtissam Bakkouri $ ^{1} $ · Siham Bakkouri $ ^{2,3} $ Abstract: Accurate segmentation of colon polyps in endoscopic images is crucial for early colorectal cancer diagnosis and treatment planning. However, achieving this is particularly challenging due to the diverse characteristics of polyps, including variations in size, color, shape, position, boundary ambiguity, and complex structure. To address these challenges, this paper introduces the Multi-level Multi-scale Gated Attentional Squeezed Network (2MGAS-Net), a robust deep learning model designed specifically for polyp segmentation. 2MGAS-Net incorporates a novel modular Multi-scale Gated Attentional Squeezed Feature Fusion (MGAS2F) strategy. MGAS2F effectively captures contextual information at multiple scales through a combination of Multi-scale Squeezed Feature Fusion (MS2F) and Cascaded Gated Attentional Transformer (CGA-T) modules. MS2F enhances the model's ability to extract detailed polyp features, while CGA-T guides the model for accurate polyp boundary estimation. Experiments on publicly available datasets demonstrate that 2MGAS-Net outperforms existing state-of-the-art methods. This indicates its potential to improve polyp segmentation accuracy significantly, facilitating more accurate clinical decision-making and potentially revolutionizing diagnostic approaches for colorectal cancer. Keywords: Polyp segmentation · Squeezed features · Multi-scale features · Cascaded gated attention · Multi-level network · Feature fusion This paper proposes a novel CNN-based algorithm specifically designed for accurate polyp segmentation in endoscopic images. The key contributions of this study can be summarized as follows: (1) It introduces a new framework called 2MGAS-Net, which integrates fused MGAS2F units for effective polyp segmentation with a high focus on small and crucial lesion boundaries. (2) It proposes an effective MS2F paradigm to capture more specific and efficient contextual information. (3) It designs a novel CGA-T module that contributes to improving the performance of our segmentation system by focusing the MGAS2F model on polyp lesions in colonoscopy images. (4) It presents a series of comparative experiments to validate the effectiveness of 2MGAS-Net. The proposed framework has been validated on publicly available endoscopic datasets and compared with state-of-the-art CNNs that have been previously evaluated in the literature. Unlike existing research that often relies on cross-level feature aggregation approach, this study introduces a novel deep learning architecture for polyp lesion segmentation. This architecture leverages a multiscale feature fusion strategy incorporating gated attention and squeezed feature fusion for improved segmentation accuracy. The performance evaluation analysis demonstrates the potential clinical value of the proposed framework.2MGAS-Net: A multi-level multi-scale gated attentional squeezed network for polyp segmentation Ibtissam Bakkouri $ ^{1} $ · Siham Bakkouri $ ^{2,3} $ Abstract: Accurate segmentation of colon polyps in endoscopic images is crucial for early colorectal cancer diagnosis and treatment planning. However, achieving this is particularly challenging due to the diverse characteristics of polyps, including variations in size, color, shape, position, boundary ambiguity, and complex structure. To address these challenges, this paper introduces the Multi-level Multi-scale Gated Attentional Squeezed Network (2MGAS-Net), a robust deep learning model designed specifically for polyp segmentation. 2MGAS-Net incorporates a novel modular Multi-scale Gated Attentional Squeezed Feature Fusion (MGAS2F) strategy. MGAS2F effectively captures contextual information at multiple scales through a combination of Multi-scale Squeezed Feature Fusion (MS2F) and Cascaded Gated Attentional Transformer (CGA-T) modules. MS2F enhances the model's ability to extract detailed polyp features, while CGA-T guides the model for accurate polyp boundary estimation. Experiments on publicly available datasets demonstrate that 2MGAS-Net outperforms existing state-of-the-art methods. This indicates its potential to improve polyp segmentation accuracy significantly, facilitating more accurate clinical decision-making and potentially revolutionizing diagnostic approaches for colorectal cancer. Keywords: Polyp segmentation · Squeezed features · Multi-scale features · Cascaded gated attention · Multi-level network · Feature fusion This paper proposes a novel CNN-based algorithm specifically designed for accurate polyp segmentation in endoscopic images. The key contributions of this study can be summarized as follows: (1) It introduces a new framework called 2MGAS-Net, which integrates fused MGAS2F units for effective polyp segmentation with a high focus on small and crucial lesion boundaries. (2) It proposes an effective MS2F paradigm to capture more specific and efficient contextual information. (3) It designs a novel CGA-T module that contributes to improving the performance of our segmentation system by focusing the MGAS2F model on polyp lesions in colonoscopy images. (4) It presents a series of comparative experiments to validate the effectiveness of 2MGAS-Net. The proposed framework has been validated on publicly available endoscopic datasets and compared with state-of-the-art CNNs that have been previously evaluated in the literature. Unlike existing research that often relies on cross-level feature aggregation approach, this study introduces a novel deep learning architecture for polyp lesion segmentation. This architecture leverages a multiscale feature fusion strategy incorporating gated attention and squeezed feature fusion for improved segmentation accuracy. The performance evaluation analysis demonstrates the potential clinical value of the proposed framework.