Seg-lite artificial intelligence: Scalable, explainable deep learning for ovarian cancer histopathology on low-resource systems

Histopathological examination remains essential for ovarian cancer diagnosis, yet routine assessment is time-consuming and depends on specialist interpretation, which may limit prompt clinical decision-making. This study proposes a compact deep-learning framework designed for automated ovarian cancer tissue segmentation that can function effectively on standard laptops and graphics processing units with limited VRAM. The model, a streamlined U-Net containing 237,457 parameters, was trained on 43,265 image patches at 64 × 64 resolution using a 70/15/15 train/validation/test split. Despite its small footprint, it achieved a Dice score, intersection-over-union, and pixel accuracy of 1.000 on the test set, converging in approximately 20 min over six epochs. Gradient-weighted class activation mapping visualization confirmed that the network consistently focused on morphologically relevant structures. These findings demonstrate that high-precision segmentation can be achieved without computationally expensive architectures, providing a practical and scalable solution for resource-limited clinical environments.