<p>Medical image segmentation is vital in the clinical field because it enhances tumour detection, outlines organs, and enables precise diagnosis, which is crucial for planning and performing surgery. However, achieving accuracy in segmentation proves challenging due to the complex contours, surface features, and magnitudes found in medical images. Techniques such as U-Net, DeepLabV3, and others have not sufficiently captured either local or global features, resulting in poorer performance in advanced tasks like tumor segmentation and organ delineation. To address these issues, this study presents a novel method called TAT-HHO, based on the Bio-Inspired Harris Hawks Optimization Tri-Attribute T-Net. This approach integrates T-Net architecture with an optimization algorithm based on Harris hawks’ hunting behavior. This hybrid approach improves the iterative process carried out for the T-blocks incorporating standard and dilated convolutional layers, capturing both local and global features, and thus refining the global and local segmentation. This model developed in this work achieved an accuracy of 99.93%, a Dice score of 99.56%, a score of 98.98% IoU, which is substantially higher than all the others in the comparison set. Also, the cross-validation in this case shows low variance, skewness, and kurtosis which reaffirms the robustness and the reliability of the model. TAT-HHO is further validated from the model with additional analysis including statistical tests and reported a value of <i>p</i> = 0.0000157 which supports the claims of enhancement made with the use of TAT-HHO. Overall, this illustrates that the framework TAT-HHO redefines the medical image segmentation paradigm by providing a highly precise, multistage, low drift and reliable framework. This is expected to increase the use of the technology in actual clinical work for the segmentation of tumors and organs.</p>