Sensitive and precise segmentation of the stomach, small bowel, and large bowel is fundamental to enhancing the accuracy of diagnosis, optimizing treatment strategies, and facilitating surgical procedures. The variability and complexity of the GI tract’s anatomy pose challenges to traditional segmentation approaches. In this work, we present the TATIMPA (Tri Attribute T-Net Integrated Multi-Pyramidal Attention) network, an advanced encoder-decoder deep learning system with multi-scale attention and residual learning tailored for organ segmentation. The proposed model features tri-attribute feature extraction, dilated convolutions, and a novel hybrid loss function, Banerjee’s Coefficient, which improves segmentation boundary and morphological coherence. With the model trained and validated on a specially curated medical dataset of annotated gastrointestinal images, TATIMPA achieved the following Dice Coefficient scores during training: 0.9932 for the stomach, 0.9952 for the small bowel, and 0.9952 for the large bowel. The model’s testing and validation scores remained just as high, with recall values over 0.99 in all phases. The model also surpassed the most advanced U-Net, ER-Net, and Res-UNet based implementations on all the most important benchmarking metrics Dice and Jaccard Index and F2 Score. Models were tested for robustness and consistency using ANOVA, Friedman, and Bland–Altman tests, confirming generalizability and independence to the specific sample population dataset. This positions TATIMPA as a potent candidate for clinical use in automated diagnosis, preoperative strategizing, as well as in longitudinal patient monitoring, precision, and reliability in gastrointestinal healthcare optimization.