<scp>TAU‐EffNetB7</scp> : A Novel Triple Attention U‐Net Approach Using <scp>EfficientNetB7</scp> for Enhanced Polyp Segmentation

ABSTRACT Polyp segmentation is a critical but challenging process in clinical imaging since colonoscopic images are inherently complex and heterogeneous. Conventional single‐stage segmentation networks lack good generalization and achieve only acceptable accuracy, particularly for small or uncertain polyps. To address these constraints, we propose two new models: TAU‐EffNetB7 and TAU‐EffNetB7 + Residual. These models apply triple‐attention U‐Net and triple‐attention residual architectures, respectively, and incorporate cascaded stages, attention and residual operations, Atrous Spatial Pyramid Pooling, and transfer learning from EfficientNetB7. The multi‐stage architecture enables progressive refinement of segmentations, better capture of multi‐scale features, and accurate depiction of intricate boundaries. We evaluate our models on three publicly available colonoscopic datasets: Kvasir‐SEG, CVC‐ClinicDB, and CVC‐ColonDB. The TAU‐EffNetB7 attains Dice Similarity Coefficients (DSC) of 89.54%, 94.62%, and 94.68% on each dataset, respectively. The TAU‐EffNetB7 + Residual model performs even better, achieving DSCs of 91.11%, 93.74%, and 94.72%, significantly outperforming baseline models such as U‐Net and Attention U‐Net. To assess generalization, we carry out experiments where models are trained with small subsets of data (Kvasir‐SEG1, CVC‐ClinicDB1, and CVC‐ColonDB1) and tested on the full datasets. Both models demonstrate strong performance even with limited training data. TAU‐EffNetB7 achieves 90.18% DSC when trained on Kvasir‐SEG1, whereas TAU‐EffNetB7 + Residual achieves 94.17% on CVC‐ClinicDB and 94.68% on CVC‐ColonDB when trained on their respective subsets. Notably, the residual‐augmented model outperforms its counterpart in all but a few low‐data scenarios.