Page 111 - AIH-2-2

Artificial Intelligence in Health AI in early breast cancer diagnosis: A review

3. Systematic literature review a lack of focus on early detection methods, classification
as questionnaires, review articles, or meta-analyses rather
3.1. Methods than primary research.
This literature review was drafted according to the
Preferred Reporting Items for Systematic Reviews and 3.2. Results
Meta-Analyses (PRISMA) guidelines. A digital search was Out of 26 studies finalized, 13 articles focused on early
conducted using standard terms and keywords in PubMed, detection of breast cancer by implementing AI in imaging
along with the application of AND/OR operators for highly modalities. The remaining 13 studies were focused on
specific results, ACM Digital Library, and SciSpace. The cancer biomarkers obtained through LB, which have the
search spanned from February to April 2024, and the query potential to revolutionize early diagnosis of breast cancer.
results were saved to Zotero. Duplicates among the saved
searches were manually scanned and removed. In total, 3.2.1. Early detection of breast cancer by AI
192 articles were identified and assessed for language, title, implementation in computer-aided detection
abstract, date of publication, and type of work. Full-text Tables 3 and 4 provide a comprehensive overview of studies
screening further narrowed down the results for exclusion, focused on implementing AI models with CAD to improve
and the remaining studies were screened for full-text the diagnostic accuracy of breast cancer.
access. This step was performed by a single reviewer.
Thirteen articles investigated the use of AI with CAD
This study adheres to the PRISMA 2020 statement. to improve the diagnostic accuracy of breast cancer across
Exclusion/inclusion was based on the criteria defined various medical imaging modalities such as ultrasound,
below. mammography, and histopathology images. Nine articles

Studies included in this study must involve human implemented CAD systems with DL models and four
participants, specifically focusing on the female population reports utilized ML models to establish a diagnostic
within the age range of 18 – 65 years or older. Eligible model to assist clinicians. Table 5 expands on the model
studies included those with patients diagnosed with breast performance metrics developed in each study.
cancer confirmed by medical examination and healthy Mahmood et al. used a novel approach to detect
20
controls with no history of cancer. Inclusion criteria mitosis in histopathology images by designing a
require clear reporting of sensitivity, total number of cases, combination of Resnet-50 and a dense convolutional
and a description of case derivation, with blood samples network (Densenet)-201, achieving a precision of 87.6%.
collected for analysis before any treatment. Comprehensive The datasets used were the International Conference
data extraction was mandated, including patient ethnicity, on Pattern Recognition (ICPR) 2012, with 50 high-
detection methods, and accurate diagnostic figures (true power field images from five different slides, ICPR 2014,
positives, false positives, false negatives, and true negatives). with 1,696 frames from 73 patients, and the Tumor
Exclusion criteria encompassed duplicate publications, Proliferation Assessment Challenge 16, which comprised
articles with insufficient data, meeting abstracts, review 821 image regions from 73 breast cancer cases. Jignesh
articles, meta-analyses, and questionnaires or knowledge Chowdary et al. worked with ultrasound images to
17
surveys. develop systematized segmentation and classification
In summary (Table 2), out of the 192 articles identified of breast cancer tumors. The Breast Ultrasound Images
in the databases, 26 studies remained for full reading after Dataset was used, which included 780 breast ultrasound
removing duplicates and articles whose titles and abstracts images (133 normal, 487 benign, and 210 malignant). The
did not meet the eligibility criteria. The analysis of this study proposed network leveraged the processing capabilities of
was based on the full-text articles, not just the abstracts of residual and U-Net architectures, allowing it to perform
the included articles, as shown in Figure 3. Reasons for segmentation and classification simultaneously. This DL
16
exclusion included irrelevance to breast cancer detection, approach achieved 97.86% accuracy and 98.12% precision
for breast cancer diagnosis. 17
Table 2. Databases used and number of articles found
Rehman et al. put forth a novel architectural
18
Databases No. of articles found No. of selected articles distortion (AD)-based mammogram classification that
PubMed 159 20 (12.6%) utilizes depth-wise DL techniques and computer vision to
classify benign and malignant breast cancer from digital
ACM Digital Library 23 6 (26.08%) mammograms. This proposed method achieved an average
SciSpace 10 1 (10%) accuracy of 97% on PINUM Cancer Hospital, Curated
Total 192 26 Breast Imaging Subset of Digital Database for Screening

Volume 2 Issue 2 (2025) 105 doi: 10.36922/aih.4197