Table 3 Summary of AI-based prognostic algorithms in breast cancer pathology
From: Artificial intelligence’s impact on breast cancer pathology: a literature review
Year of publication | Reference | Number of slides for training / validation | Pathologists review (training/validation) | Algorithm details | Algorithm endpoints/outputs | Algorithm performance |
|---|---|---|---|---|---|---|
2020 | Pantanowitz et al [24] | 320 breast invasive ductal carcinoma cases, 16,800 digital image patches from 120 WSIs/140 digital image | Ten expert pathologists and 24 readers of varying expertise | RCNN by ResNet-101 | Mitotic figures counting | Accuracy with AI = 55.2% compared to manually 43.9% |
2020 | Chow et al [25] | 93 cases of phyllodes tumor | N/A | Image Management System viewer | Phyllodes tumor mitoses counting | correlation = .794; R2 = 0.63; P < .001; 95% CI, 0.270–0.373 |
2021 | Balkenhol et al [26] | 94 TNBC specimens | two histopathologists | convolutional neural networks (CNN) | TILs assessment and prognostic values | Relapse free survival HR ranging between 0.777 (CD8, IM2) and 0.915 (CD3, ITS); overall survival HR varying between 0.722 (FOXP3, ITT) and 0.908 (CD3, ITA) |
2022 | Wang et al [6] | Training:1567Test:1262 | Pathologists | deep CNN model | Categorization of NHG2 breast tumors and its risk of recurrence | increased risk for recurrence in DG2-high (HR 1.91, 95% CI 1.11–3.29, P = 0.019) |
2022 | Mantrala et al [27] | Training: 46 Test: 91 | Six pathologists | Unet, DenseNet backbone, preloaded ImageNet for TF, HoVerNet, pretrained ImageNet ResNet50-Preact for NP, LinkNet with EfficientNet B4 backbone for MC | Breast cancer grading | Tubular formation (κ = 0.471 each) Nuclear pleomorphism (κ = 0.342) and was worst for mitotic count (κ = 0.233) |