Enhancing explainability in pacu fish image segmentation using saliency maps and combined explainable AI methods

Advances in Artificial Intelligence (AI) have sparked concerns regarding the transparency of model outputs, necessitating the development of eXplainable Artificial Intelligence (XAI) techniques. This study presents an improved evaluation of XAI methods applied to Pacu fish image segmentation. We compare and evaluate four XAI methods - Grad-CAM, Saliency Map, CNN Filters, and Layer Grad-CAM - using pixel perturbation techniques applied in 100 fish images. Our experiments reveal that through the images generated by the pixel perturbation techniques (118 to the white noise, 144 to the dark noise, and 123 to the random noise), the Saliency Map achieves the best results, highlighting the most relevant regions for AI model predictions. Furthermore, by combining Saliency Maps with other XAI methods, we demonstrate substantial improvements in explainability and segmentation accuracy. In this case, the largest number of images generated in the second experiment was 108, by combining the between Saliency Map and Grad-CAM to the white noise pixel perturbation. Finally, this work not only advances the state-of-the-art in XAI for image segmentation but also underscores the importance of combining XAI techniques to achieve superior explanatory power in areas such as Aquaculture. • Evaluating XAI methods using pixel perturbation is a correct approach supported by the literature. • Saliency Map had the best performance, revealing the most relevant pixels according to Mask R-CNN. • The combination of the Saliency Map served as a determining factor for the improvement of the three other XAI methods tested. • For the aquaculture sector, improving the explainability of XAI methods avoids erroneous and misleading.