BI-RADS-Derived Breast Cancer Risk Classification from Mammograms

Mammography is a central imaging modality for breast-cancer screening, but interpretation remains difficult because suspicious findings can be subtle and breast tissue can be heterogeneous or dense. This study evaluates whether interpretable, hand-crafted digital image-processing features can support BI-RADS-derived low-risk versus high-risk classification using the INbreast full-field digital mammography dataset, which contained 410 mammograms from 108 anonymized patients. Labels were derived from BI-RADS categories (1-3: low risk; 4-6: high risk), and BI-RADS 0 examinations were excluded from the assigned modeling sets. A leakage-aware patient-level split produced 240 training, 84 validation, and 86 test images. The image-processing pipeline explored gray-level enhancement, convolution and filtering, binary-mask morphology, segmentation and region-of-interest construction, and then extracted 60 interpretable features spanning intensity statistics, histogram distribution, gray-level co-occurrence matrix texture, histogram of oriented gradients, and shape. Classical classifiers were compared using raw and train-fitted min-max-normalized features, with feature subsets selected using t-test-based incremental feature selection (tIFS) and Random-Forest recursive feature elimination (RF-RFE). On the held-out test set, the highest AUROC among evaluated final configurations was 0.7382 for RF-RFE Random Forest with min-max normalization; however, sensitivity was only 0.0417, demonstrating that overall discrimination or accuracy alone can obscure poor detection of the minority high-risk class. A more balanced RF-RFE SVM achieved AUROC 0.6969, sensitivity 0.5417, specificity 0.7581, and MCC 0.2869. These findings support interpretable image features as a transparent baseline while showing limitations from class imbalance, small sample size, and feature representation. The study represents BI-RADS-derived risk classification rather than pathology-confirmed cancer diagnosis.

Authors

Publication Details

Journal
Theoretical and Natural Science
Published
2026-11-17
DOI
https://doi.org/10.54254/2753-8818/2026.hz38953
Primary Topic
AI in cancer detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

BI-RADS-Derived Breast Cancer Risk Classification from Mammograms

Zihan Wu
1 links
Theoretical and Natural Science
AI in cancer detection
article

BI-RADS-Derived Breast Cancer Risk Classification from Mammograms

Zihan Wu
article en

Abstract

Mammography is a central imaging modality for breast-cancer screening, but interpretation remains difficult because suspicious findings can be subtle and breast tissue can be heterogeneous or dense. This study evaluates whether interpretable, hand-crafted digital image-processing features can support BI-RADS-derived low-risk versus high-risk classification using the INbreast full-field digital mammography dataset, which contained 410 mammograms from 108 anonymized patients. Labels were derived from BI-RADS categories (1-3: low risk; 4-6: high risk), and BI-RADS 0 examinations were excluded from the assigned modeling sets. A leakage-aware patient-level split produced 240 training, 84 validation, and 86 test images. The image-processing pipeline explored gray-level enhancement, convolution and filtering, binary-mask morphology, segmentation and region-of-interest construction, and then extracted 60 interpretable features spanning intensity statistics, histogram distribution, gray-level co-occurrence matrix texture, histogram of oriented gradients, and shape. Classical classifiers were compared using raw and train-fitted min-max-normalized features, with feature subsets selected using t-test-based incremental feature selection (tIFS) and Random-Forest recursive feature elimination (RF-RFE). On the held-out test set, the highest AUROC among evaluated final configurations was 0.7382 for RF-RFE Random Forest with min-max normalization; however, sensitivity was only 0.0417, demonstrating that overall discrimination or accuracy alone can obscure poor detection of the minority high-risk class. A more balanced RF-RFE SVM achieved AUROC 0.6969, sensitivity 0.5417, specificity 0.7581, and MCC 0.2869. These findings support interpretable image features as a transparent baseline while showing limitations from class imbalance, small sample size, and feature representation. The study represents BI-RADS-derived risk classification rather than pathology-confirmed cancer diagnosis.

Theoretical and Natural ScienceVol. 191(1)
Good health and well-being
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon