[Paper] Scalable Residual Feature Aggregation Framework with Hybrid Metaheuristic Optimization for Robust Early Pancreatic Neoplasm Detection in Multimodal CT Imaging

Source: arXiv - 2512.23597v1

Overview

Early detection of pancreatic neoplasms remains a tough clinical problem because tumors often appear with faint contrast and highly variable anatomy on CT scans. The authors present a Scalable Residual Feature Aggregation (SRFA) framework that stitches together modern deep‑learning blocks, a hybrid meta‑heuristic feature selector, and dual‑stage hyper‑parameter optimization to boost both accuracy and robustness on multimodal CT data.

Key Contributions

• SRFA pipeline that combines preprocessing, a custom MAGRes‑UNet segmentation head, and residual feature aggregation via DenseNet‑121.
• Hybrid meta‑heuristic feature selection (Harris Hawks Optimization + Bat Algorithm) to prune the high‑dimensional feature space without sacrificing discriminative power.
• Hybrid classifier that fuses Vision Transformer (ViT) attention mechanisms with EfficientNet‑B3’s efficient representation learning.
• Dual optimization of hyper‑parameters using Salp Swarm Algorithm (SSA) and Grey Wolf Optimizer (GWO) to curb over‑fitting and improve generalization.
• State‑of‑the‑art performance on a multi‑modal pancreatic CT dataset: 96.23 % accuracy, 95.58 % F1‑score, 94.83 % specificity, outperforming standard CNNs and recent transformer‑based baselines.

Methodology

1. Pre‑processing & Segmentation – Raw CT volumes are normalized and fed into MAGRes‑UNet, a UNet variant enriched with residual connections that isolates the pancreas and surrounding tissue, sharpening subtle lesion cues.
2. Feature Extraction & Aggregation – The segmented ROI is passed through DenseNet‑121 equipped with a residual feature storage layer. This preserves hierarchical features across layers, allowing the network to retain fine‑grained texture information that typical pooling would discard.
3. Meta‑heuristic Feature Selection – The high‑dimensional feature map is trimmed using a hybrid Harris Hawks Optimization (HHO) + Bat Algorithm (BA) search. The algorithm iteratively evaluates subsets against a fitness function (classification accuracy vs. feature count) to keep only the most informative descriptors.
4. Hybrid Classification Model – Selected features are fed into a Vision Transformer (ViT) that captures global contextual relationships, while an EfficientNet‑B3 branch supplies compact, high‑capacity embeddings. The two streams are concatenated and passed through a final fully‑connected head.
5. Hyper‑parameter Fine‑tuning – A two‑phase optimizer—first Salp Swarm Algorithm (SSA), then Grey Wolf Optimizer (GWO)—searches the space of learning rates, weight decay, and transformer depth, automatically balancing exploration and exploitation to reduce over‑fitting.

Results & Findings

• Classification Accuracy: 96.23 % (vs. ~88 % for baseline ResNet‑50).
• F1‑Score: 95.58 %, indicating strong balance between precision and recall on a highly imbalanced medical dataset.
• Specificity: 94.83 %, crucial for minimizing false positives in a screening context.
• Ablation studies show each component (MAGRes‑UNet, hybrid feature selector, dual optimizer) contributes ~2–4 % incremental gains, confirming the synergy of the pipeline.
• Training stability improves markedly; loss curves plateau earlier and exhibit lower variance, reflecting the effectiveness of the dual hyper‑parameter optimizer.

Practical Implications

• Clinical Decision Support: Radiology teams can integrate the SRFA model into PACS workstations to flag suspicious pancreatic regions automatically, reducing missed early lesions.
• Scalable Deployment: The modular design (segmentation → feature extraction → selection → classification) maps cleanly onto micro‑service architectures, allowing cloud‑based inference pipelines that can scale with hospital imaging volumes.
• Developer Friendly: All core components rely on widely‑used libraries (PyTorch, TensorFlow, OpenCV) and the meta‑heuristic optimizers are implemented as lightweight Python modules, easing reproducibility and customization for other organ‑level detection tasks.
• Transferability: Because the framework is modality‑agnostic, developers can repurpose it for other low‑contrast CT or MRI problems (e.g., early liver tumor detection) by swapping the segmentation head and fine‑tuning the optimizer settings.

Limitations & Future Work

• Dataset Diversity: The study uses a single institutional CT cohort; broader multi‑center validation is needed to confirm robustness across scanner models and acquisition protocols.
• Computation Overhead: The hybrid meta‑heuristic search and dual optimizer add noticeable pre‑training time, which may be prohibitive for rapid prototyping without GPU clusters.
• Explainability: While the ViT attention maps provide some visual insight, the paper does not quantify interpretability metrics—future work could integrate saliency or SHAP analyses for clinician trust.
• Real‑World Integration: Prospective trials measuring impact on diagnostic workflow, false‑positive rates in routine screening, and cost‑benefit analyses remain open research avenues.

---

Bottom line: The SRFA framework showcases how a thoughtfully engineered blend of deep‑learning architectures and bio‑inspired optimization can push the envelope for early pancreatic cancer detection—offering developers a concrete, extensible blueprint for tackling other subtle‑signal medical imaging challenges.

Authors

• Janani Annur Thiruvengadam
• Kiran Mayee Nabigaru
• Anusha Kovi

Paper Information

• arXiv ID: 2512.23597v1
• Categories: cs.CV, cs.IR
• Published: December 29, 2025
• PDF: Download PDF