A newly published study details a reproducible breast mass segmentation pipeline designed to tackle domain shift challenges in computer-aided diagnosis using an improved crested porcupine optimizer.
Researchers developed a strong U-Net segmentation model optimized through a novel algorithm instead of traditional gradient-based methods. This approach is engineered to overcome non-convex loss surfaces and avoid local minima during mammogram analysis.
The framework relies on rigorous preprocessing applied consistently across both the INbreast and CBIS-DDSM datasets. Patient-wise splitting and uniform 256 by 256 region of interest extraction form the bedrock of the pipeline.
A two-phase training strategy utilizes a composite loss function optimized with the Improved Crested Porcupine Optimizer, or ICPO. Validation curves demonstrate that the networks converge cleanly without catastrophic overfitting across heterogeneous clinical domains.
Performance metrics show substantial gains over baseline optimization techniques. Across three specific case studies—INbreast, CBIS-DDSM, and mixed training—the framework achieved high accuracy ratings.
Dice Similarity Coefficient values reached 0.8721, 0.9623, and 0.9187 respectively. IoU values hit 0.7704, 0.9289, and 0.8452 across the same evaluation sets.
These outcomes improve upon the existing state-of-the-art AdamW optimizer baseline by 0.9 to 2.3 percent. Generalizability across heterogeneous mammography domains is verified by explicit random seed controls, published manifest files, and freely accessible methodologies.
While the study demonstrates convergence and cross-domain robustness on benchmark datasets, questions remain regarding performance on unseen clinical scanners with varying acquisition parameters. Future deployments must prove whether the ICPO-optimized U-Net pipeline maintains its 0.9 to 2.3 percent advantage.
