Abstract
Pancreatic cancer remains one of the deadliest cancers due to the lack of effective early detection tools. While deep neural networks (DNNs) have shown promise in tumor segmentation, electronic accelerators suffer from power inefficiency and latency. To address this, we propose MediONN-a photonic neural network system implemented on an integrated chip, optimized for 3D medical image segmentation. MediONN integrates a 4×4 photonic neural processor within a hierarchical 3D optical computation framework. To improve training convergence, we introduce a segmentation-specific Gaussian weight initialization strategy, along with 3D optical convolutional layers for volumetric feature extraction. Unlike prior photonic systems focused on classification, MediONN is the first to demonstrate optical neural networks (ONNs) directly applied to 3D segmentation. On the NIH pancreas CT dataset, MediONN achieves a Dice similarity coefficient (DSC) of 0.5215 (2D) and 0.5302 (3D), with peak DSCs of 0.5919 (2D) and 0.8788 (3D). Comprehensive evaluation metrics confirm MediONN's segmentation accuracy is comparable to electronic counterparts, while offering significant gains in computational speed and energy efficiency. These results highlight the scalability and biomedical potential of integrated photonic ONNs.