Table 3.
AI Applications in Breast Cancer Diagnostics (2018–2024). Representative Studies Across Imaging and Biosensing, Summarizing Datasets, Model Families, Headline Performance, and Key Findings
| Ref. | Year | Modality | Dataset | AI Model/Technique | Overall Accuracy/(AUC-ROC) | Key Findings |
|---|---|---|---|---|---|---|
| [16] | 2020 | Mammography & Digital Breast Tomosynthesis (DBT) | Large representative dataset (UK), large enriched dataset (USA) | DeepMind AI | 11.5% AUC | Improved sensitivity and specificity over radiologists (+11.5%, +5.7%) |
| [91] | 2020 | Screening Mammography | 240 women (100 cancers, 40 false positives, 100 normal; 2013–2017) | AI support with decision support and lesion markers | 0.89% AUC | AI support improved AUC from 0.87 to 0.89; increased sensitivity (86% vs 83%); specificity |
| [92] | 2018 | Ultrasound Imaging | Dataset A: 306 images (60 malignant, 246 benign) -Dataset B: 163 images (53 malignant, 110 benign) | Patch-based LeNet - U-Net - Transfer learning (pretrained FCN-AlexNet) | Not specified | Deep learning outperforms state-of-the-art, with improved metrics and Dataset B available. |
| [93] | 2021 | Ultrasound Imaging | Breast Ultrasound Images (BUSI) - Mendeley Breast Ultrasound Dataset - Dataset A: 306 images - Dataset B: 163 images | Transfer Learning (various CNNs) | 99% accuracy | Achieved accuracies: 97.8% (BUSI), 99% (Mendeley), 98.7% (MT-Small). |
| [94] | 2022 | Multiparametric MRI | 93 women, 104 histopathologically verified lesions | Radiomics + Machine Learning (10 predictive models) | AUC = 0.93–0.96 | Radiomics with DWI and ADC boosts accuracy (88.5%) over radiologists (85.6%), aiding less experienced readers. |
| [95] | 2023 | Breast MRI | Multiple datasets | CNN, RNN, RCNN, AE, Ensemble models | AUC-ROC: up to 0.98 | Deep learning aided in breast cancer diagnosis, molecular classification, chemotherapy response, and lymph node prediction. |
| [96] | 2023 | Pathology Imaging | H&E and fluorescent-stained datasets | Color Normalization, Nucleus Extraction, Segmentation AI Model | H&E: 89.6%, Fluorescent: 80.5% | Cross-staining inference achieved high accuracy, demonstrating potential for expanding current pathology AI models to different staining technologies. |
| [97] | 2023 | Histopathological Imaging | 95 TNBC and HER2+ BC H&E-stained WSI (1037 regions of interest | Pathomic feature extraction, Random-Forest, other ML models | AUC-ROC: 0.86 | Enhanced consistency in tumor grading |
| [98] | 2024 | Electrochemical Biosensors | - | Signal Processing AI | 90.0 | Improved biomarker detection for early diagnosis |
Notes: Values are reported in the original studies. Accuracy is given in % and AUC-ROC on a 0–1 scale. “NR” = not reported; “pp” = percentage-point difference versus the comparator. Where both baseline and post-AI AUC are available, the change is shown (eg, 0.87 → 0.89).
Abbreviations: AE, autoencoder; ADC, apparent diffusion coefficient; AUC-ROC, area under the receiver operating characteristic curve; B, benign; BUSI, Breast Ultrasound Images dataset; DBT, digital breast tomosynthesis; DL, deep learning; DWI, diffusion-weighted imaging; FP, false positive; H&E, hematoxylin and eosin; M, malignant; ML, machine learning; RCNN, region-based convolutional neural network; RNN, recurrent neural network; SOTA, state of the art; TL, transfer learning; TNBC, triple-negative breast cancer; HER2+, human epidermal growth factor receptor 2–positive; WSI, whole-slide image; pp, percentage points.