Page 116 - AIH-2-2

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

mammography images; and DDSM includes 2500 studies. maps between the encoder and decoder, achieving an F1
However, improvements are needed in recall rates and score of 0.81. However, the complexity of the network
capturing AD tracks from digital breast tomosynthesis architecture and its testing solely on private breast tumor
slices. Jianwei Liao et al. developed a DL model to detect ultrasound images may limit its performance on other
19
high-risk sites in ultrasound images that appear benign, tumor types. In addition, Wang et al. fused biological and
24
facilitating early cancer confirmation with urgent biopsy. The computational methods to detect early-stage breast cancer
reported accuracy was 94.4% on private datasets obtained using Raman spectroscopy. Raman spectra were collected
from three different hospitals in China, outperforming from 804 serum samples, of which 241 were from healthy
radiologists. The training dataset had Doppler ultrasound volunteers, 463 from breast cancer patients, and 100 from
images of 7955 lesions from 6795 patients. The model’s ductal carcinoma in situ patients – provided by Peking
performance was validated in two external cohorts that University People’s Hospital. The work combined Raman
included 448 lesions from 391 patients in the Tangshan spectroscopy of the serum with CNNs, demonstrating
People’s Hospital, Chongqing, China, and 245 lesions from high capability for rapid, accurate, and minimally invasive
235 patients in the Dazu People’s Hospital, Chongqing, breast cancer screening.
China. The system was designed to assist radiologists Fogliatto et al. applied ML techniques to the Wisconsin
21
and the experimental results indicated that this ensemble Breast Cancer Database (WBCD) and a protein microarray
DL model could enhance the diagnostic efficiency of database to design a decision support system for early
radiologists by identifying subtle and revealing elements breast cancer detection. The WBCD consists of 699
on ultrasound images of breast tissues. Another ensemble
DL model studied by Subramanian and Venugopal instances with nine attributes, and the protein microarray
26
achieved an accuracy of 96.71% through hyperparameter database includes 642 features regarding the levels of
optimization. The model was trained on the Breast Cancer protein expression analyzed by microarray chips. The study
Wisconsin Diagnostic dataset containing 569 instances, ranked features using a new feature importance index based
the Breast Cancer Wisconsin Original dataset with 699 on principal component analysis and Bayesian decision
instances, and the Breast Cancer Coimbra dataset with 116 parameters, yielding an average accuracy of 98.3% with
instances. While this model performed better than existing k-nearest neighbors (kNN), linear discriminant analysis
ones in the literature, it was validated on the same dataset; (LDA), and probabilistic neural network (PNN). Nomani
23
it was trained on rather than an external study, raising et al. attempted to improve early breast cancer diagnosis
23
concerns about its applicability in clinical settings. by image processing of mammograms from the Mini-
MIAS dataset, which contains 322 digitized mammograms.
Ayana et al. designed a novel patchless multi-stage The study developed a novel particle swarm-optimized
22
transfer learning technique to improve breast mass wavelet neural network that achieved a precision of 98.6%
classification. The datasets used were DDSM with 2620 on screening MIAS data, outperforming SVM, kNN, and
22
mammography images, INbreast database with 115 cases CNN. Malek et al. presented a new methodology to
28
(410 images), and Mammographic Image Analysis Society classify microcalcifications in mammogram images using
(MIAS) comprising 322 digitized film mammograms. persistent homology (PH) and ML models. The study
28
The model outperformed patch-and whole image- used MIAS and DDSM datasets containing 322 digitized
based methods, achieving an improved accuracy by 8% film mammograms and 2620 mammography images,
(91.41% vs. 99.34%) when tested on the INbreast dataset. respectively. The study demonstrated the use of PH in
Despite the high accuracy, further fine-tuning could be noise filtration and feature extraction, the results of which
beneficial if patient information and medical history are were used to train ML models. This approach achieved an
also considered. Yasir et al. implemented YOLOv5 in accuracy of 96.2% and 99.3% on MIAS and DDSM datasets,
27
combination with Mask R-convolutional neural networks leading to the discovery that PH-based ML methods can
(CNNs) on the INbreast dataset, achieving an accuracy increase classification performance in terms of accuracy. PH
of 98%. This novel technique proved to be better than a can capture topological features that conventional image-
single detection unit; however, it was conducted on a single analysis methods cannot extract, paving the way for a more
dataset and not validated against others. accurate diagnosis. Alfian et al. developed a web-based
29
Another study by Yu et al. focused on tumor application predicting breast cancer based on underlying
25
segmentation of breast ultrasound images, by enhancing risk factors with an accuracy of 80.23% using the Coimbra
the original U-Net model. The study used a private dataset dataset comprising 116 breast cancer instances. Although
from Xinhua Hospital, which contained 538 tumor this model was designed to assist healthcare professionals
images. The proposed approach integrated a Res Path using decision trees and SVM classifiers, its accuracy may
into the U-Net model to reduce discrepancies in feature not be sufficient for clinical reliance.

Volume 2 Issue 2 (2025) 110 doi: 10.36922/aih.4197