Page 113 - AIH-2-2

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

Table 3. (Continued)
Technique (s) References Year Approach(es) Inference Limitations
employed
Feature-based 20 2020 Faster R-CNN and This proposed technique can • Need to further improve the
deep CNNs for potentially develop an automated accuracies and lower the
the object (mitotic cancer grading system to serve as computational cost of the technique
cells) detection a second opinion for pathologists. by developing a customized network.
It can also flag cases that require • Validation of the technique is
immediate attention for cancer required using larger datasets
diagnosis. that represent a diversity of breast
cancer patients.
Feature-based 21 2019 Feature selection This study reported 99.17% accuracy Not mentioned.
in WBCD, retaining an average
of 4.61 of the nine features in the
dataset. For the protein microarray
database, the technique resulted in
an average accuracy of 98.30%, with
an average of 16.96 features retained
out of 642 total features.
Image-processing 22 2022 Patchless The patchless approach performed Only image information was used
multi-stage transfer 8% better in terms of accuracy than to retrieve the results of diagnosis;
learning the patch- and whole image-based the patient’s medical data were not
method (91.41% vs. 99.34%) when used, which is the norm for clinical
tested on the INbreast dataset. diagnosis.
Image-processing 23 2022 PSOWNNs-CNN PSOWNNs perform better than Not mentioned
SVM, KNN, and CNN in terms of
accuracy.
Raman 24 2023 Raman Serum Raman spectroscopy, used in Needs to be validated on diverse
spectroscopy spectroscopy of conjunction with CNN, has shown samples.
serum combined great potential for non-invasive and
with CNN early diagnosis of breast cancer.
Region-based 25 2022 Res path combined This study introduces an improved Complex network, which is only
with dense CNN version of the original U-Net model tested on private breast tumor
ultrasound images. A small dataset of
ultrasound tumor images may lead to
degradability in performance.
Ensemble 26 2022 Deep ensemble Performs better than other reported The model was validated on the same
network deep ensemble strategies in dataset that it was trained on, and not
literature on the dataset from another study.
Image-processing 27 2022 YOLOv5 in The two-step method Tested on one dataset only.
combination with (YOLOv5+Mask RCNN) is a better
Mask RCNN and more efficient approach as
compared to a single detection
technique.
Image-processing 28 2023 Persistent The topological features extracted by Needs to be validated by other studies.
homology-based PH provide more insight that may
machine learning not be gained by conventional image
method processing methods, leading to a
more accurate diagnosis.
Feature-selection 29 2022 Extra trees with - Accuracy is not very high.
SVM
Abbreviations: AI: Artificial intelligence; CNN: Convolutional neural network; DBT: Digital breast tomosynthesis; DL: Deep learning; EDL: Elastic deep
learning; K-NN: K-nearest neighbors; PH: Persistent-homology; PSOWNN: Particle swarm optimized wavelet neural network; RCNN: Region-based
convolutional neural network; RF: Random forest; WBCD: Wisconsin Breast Cancer Database; w.r.t: With respect to.

Mammography (CBIS-DDSM), and Digital Database for private dataset from PINUM Cancer Hospital, Faisalabad,
Screening Mammography (DDSM) datasets. PINUM is a Pakistan; CBIS-DDSM is a curated subset containing 2620

Volume 2 Issue 2 (2025) 107 doi: 10.36922/aih.4197