European Radiology: AI-based computer-aided diagnosis in mammography

At this stage, mammogram is the standard way
to screen and detect breast cancer.
However, due to the inherent limitations of its two-dimensional projection, the sensitivity is low, between
63% and 84%.
Currently, AI-based computer-aided diagnostics (AI-CAD) are used in mammograms and have shown performance comparable to or higher than stand-alone diagnostics, while significantly improving the diagnostic performance
of radiologists when used as an additional tool.
Unlike traditional CAD programs based on features defined by radiologists, AI-CAD algorithms based on deep learning networks are less intuitive
in deriving final scores or results.
Since most commercial AI-CAD programs provide heat maps with anomaly scores, users generally accept that a higher score means a higher
probability of cancer.
However, the significance of the scores themselves has not been clarified, and questions remain
about what different scores mean for patients, or how abnormal scores are related to the American Society of Radiology Breast Imaging and Reporting System (BI-RADS).

Recently, a study published in the journal European Radiology evaluated the abnormality score of breast cancer based on clinical, radiological and pathological features, and analyzed the breast cancer characteristics of false-negative cases in AI-CAD, providing a reference for the technology to better assist radiologists and
clinicians.

This study included 930 people diagnosed with breast cancer from January 2017 to December 2017
.
Commercial AI-CAD was applied to mammography with an abnormal score
.
Abnormal scores
were assessed based on clinical, radiological, and pathological features.
A false-negative result is defined as an abnormal score of less than 10
.

The median abnormal score for 930 breasts was 87.
4 (range 0-99).

The false negative rate of AI-CAD is 19.
4% (180/930).

Compared with low-scoring cancers, cancers with an abnormal score of more than 90 showed a high proportion of palpable lesions, BI-RADS 4c and 5 lesions, masses presenting with or without microcalcifications, and invasive breast cancer (all P < 0.
001
).
False-negative breast cancer is more likely to occur in asymptomatic patients and dense breasts than detected breast cancer, and is diagnosed with occult breast cancer and DCIS
.

Figure: A 44-year-old patient diagnosed with a 20 mm sized triple-negative subtype invasive ductal carcinoma in the left breast with a stage of T1cN0
.
The left-sided breast cancer appears as a mass (arrow) at 1 o'clock, which is missed by AI-CAD (abnormal score 7).

This study shows that high abnormal scores for breast cancer described on AI-CAD are associated
with higher BI-RADS categories, invasive pathology, and higher malignancy grades.

Original source:

Si Eun Lee,Kyunghwa Han,Jung Hyun Yoon,et al.
Depiction of breast cancers on digital mammograms by artificial intelligence-based computer-assisted diagnosis according to cancer characteristics.
DOI:10.
1007/s00330-022-08718-2