Pancreatic cancer has one of the lowest survival rates of any cancer, making early and accurate detection critical. CT scans are routinely used to image the pancreas, but manually identifying and segmenting tumors from these scans is time-consuming and requires significant expertise.
Computer-aided diagnosis (CAD) systems have emerged as powerful tools to assist radiologists. However, accurately separating cancerous tissue from normal pancreatic tissue in CT scans remains challenging because the pancreas varies greatly in shape and size and is surrounded by other organs.
The proposed system, called TSADL-PCSC (Tunicate Swarm Algorithm with Deep Learning for Pancreatic Cancer Segmentation and Classification), works in two stages. First, a W-Net model — an advanced version of the U-Net architecture with dual encoder-decoder pathways — segments the affected tumor regions from CT images.
After segmentation, a GhostNet feature extractor captures key image features efficiently, using a lightweight design that reduces computational cost. These features are then passed to an Echo State Network (ESN) classifier, which makes the final determination of whether tissue is cancerous or non-cancerous.
A key innovation in this study is the use of the Tunicate Swarm Algorithm (TSA) to tune the AI model's hyperparameters. TSA is inspired by how tunicates — marine invertebrates — use swarm intelligence and jet propulsion to hunt for prey.
The algorithm optimizes the AI model's parameters by simulating how a swarm of agents collectively move toward the best solution, avoiding collisions with each other. This approach helps the model avoid getting stuck in suboptimal solutions and improves overall classification performance.
The TSADL-PCSC system demonstrated improved segmentation accuracy over baseline approaches, with the W-Net architecture outperforming standard U-Net in delineating tumor boundaries. The TSA optimization further boosted classification performance compared to models using default hyperparameters.
The system achieved competitive accuracy in distinguishing pancreatic cancer from non-cancerous pancreatic tissue in CT images, demonstrating that the combined approach of specialized segmentation, efficient feature extraction, and nature-inspired optimization works effectively.
The TSADL-PCSC framework demonstrates that combining advanced deep learning architectures with bio-inspired optimization algorithms can yield high-performing AI systems for pancreatic cancer detection from CT scans.
Future work should validate the system on larger, multicenter CT datasets and explore real-time integration into clinical workflows. Interpretability tools would also help radiologists understand why the AI made specific decisions, building the trust needed for clinical adoption.