Related Experiment Video
Updated: Jun 27, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Hybrid Metaheuristic Feature Selection for Breast Cancer Detection in Digital Mammography: A Feasibility Study with
Bandar S Alshreef1, Yousif A Kariri1
1Department of Clinical Laboratory Sciences, College of Applied Medical Sciences, Shaqra University, Shaqra 11961, Saudi Arabia.
Journal of Clinical Medicine
|June 26, 2026
Summary
This study reassessed a mammography AI feature selection framework, finding it feasible for stress testing but limited in generalizability. External validation showed near-random performance, highlighting the need for larger datasets and rigorous validation before clinical use.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Radiomics and Deep Learning in Mammography
- Computational Pathology
Background:
- The
- small-n-large-p
- dilemma in mammography AI (many features, few cases) causes overfitting and poor generalization.
- This study evaluates the HiTopology-GOA-CSA feature-selection framework on a larger dataset with stricter evaluation.
Purpose of the Study:
- To reassess the internal performance of the HiTopology-GOA-CSA feature-selection framework for mammography AI.
- To evaluate the framework's generalizability using a larger Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) cohort and a stricter leakage-aware strategy.
Main Methods:
- Retrospective analysis of 98 CBIS-DDSM mass cases (49 benign, 49 malignant).
- Extraction of 1074 features (radiomic and deep learning), with HiTopology-GOA-CSA selecting 102 features (91% reduction).
- Comparison of two evaluation modes (standard vs. nested feature selection) using 5-fold cross-validation and a multilayer perceptron (MLP) classifier.
Main Results:
- Mode A (standard evaluation) achieved an AUC of 0.726; Mode B (nested evaluation) achieved an AUC of 0.683.
- Feature stability was moderate (Jaccard similarity 0.604).
- External validation on VinDr-Mammo (n=25) showed near-random performance (AUC 0.500) with complete prediction collapse.
Conclusions:
- The HiTopology-GOA-CSA framework is feasible for feature selection and stress testing in small-cohort mammography AI.
- Limited generalizability and prediction collapse on external validation underscore the need for larger, multicenter datasets.
- Rigorous validation, including benchmark comparisons and transparent uncertainty reporting, is crucial before clinical translation.