Author:
Mohammad Motaghianfar
Abstract
Breast cancer diagnosis with AI faces hurdles like uneven datasets, overly confident wrong predictions, and no way to measure uncertainty, which erodes trust in clinical settings. Many models don’t offer solid confidence levels, especially for tricky cases in unbalanced data.
We introduce MUVED-BC, a new AI framework that views average, error, and worst-case tumor traits as separate inputs. It uses simple parallel encoders for each view, then applies evidence-based uncertainty tools from subjective logic. We tested it on the Wisconsin Breast Cancer Dataset (569 samples) with an 80-20 split and weighted training for balance.
The model set new benchmarks: 99.12% accuracy, 100% precision, 97.62% sensitivity, and 99.77% AUC-ROC. It flagged tough cases well, with wrong predictions showing 1.7 times higher uncertainty. The top 10% uncertain cases still hit 91.67% accuracy, proving good risk sorting.
This framework offers a practical AI tool for safe diagnosis, blending high accuracy with trustworthy uncertainty measures for better doctor-AI teamwork. It tackles key safety issues in medical AI.
Keywords: Breast Cancer Diagnosis, Evidential Deep Learning, Uncertainty Quantification, Multi-View Learning, Medical AI Safety
1. Introduction
Breast cancer is a top cause of cancer deaths globally, making early, accurate detection vital for better outcomes. The Wisconsin Breast Cancer Dataset (WBCD) has helped build many AI models for cell-based diagnosis. But current methods struggle with overconfident predictions, handling uneven classes, and showing uncertainty.
Models like Support Vector Machines and basic neural networks score high but give unreliable confidence scores, often too sure when wrong. This is risky in medicine, where knowing uncertainty aids decisions. The WBCD’s mild imbalance (62.7% benign, 37.3% malignant) worsens this, lowering detection of critical malignant cases.
New evidential AI methods offer better uncertainty tracking, and multi-view learning uses different data angles. But no one has merged these for breast cancer while focusing on AI safety.
Our goal: Build and test MUVED-BC, a multi-view evidential AI for accurate breast cancer detection with solid uncertainty measures for safe use.
We predict: Mixing multi-view inputs with evidential uncertainty will keep high accuracy while spotting cases needing expert checks.
2. Methods
Data and Materials
We used the public Wisconsin Breast Cancer Dataset (WBCD) with 569 fine-needle samples and 30 features from cell images. It has 357 benign (62.7%) and 212 malignant (37.3%) cases, mirroring real imbalances.
Preprocessing and Feature Engineering
We grouped features into three views: averages (10 features), errors (10), and worst-cases (10). Each was standardized separately. Labels were binary (0: benign, 1: malignant). We used an 80-20 split keeping class ratios.
MUVED-BC Architecture
Our framework includes:
-
Multi-View Encoders: Three parallel networks (16→8 neurons) for each view.
-
Fusion Layer: Combined outputs through a 16-neuron network.
-
Evidential Output: A linear layer for evidence parameters in a Dirichlet distribution.
-
Uncertainty: Calculated as u = K/∑(evidence + 1).
The loss function mixes prediction error with uncertainty:
ℒ = ∑(y – ŷ)² + ∑(ŷ(1-ŷ)/(S+1)), where S = ∑(evidence + 1).
Training Configuration
We used Adam optimizer (lr=0.001) with learning rate reduction. Classes were weighted inversely (benign: 0.798, malignant: 1.338). We added dropout (20-30%) and early stopping (patience=10) for better generalization.
Evaluation Metrics
We measured accuracy, precision, sensitivity, specificity, AUC-ROC, and F1-score. Uncertainty was checked by comparing correct/incorrect predictions and high-uncertainty accuracy. We used 5-fold cross-validation and compared to SVM, Random Forest, and basic neural networks.
3. Results
MUVED-BC excelled: 99.12% accuracy, 100% precision, 97.62% sensitivity, 100% specificity, 99.77% AUC-ROC, and 98.80% F1-score. It beat benchmarks like SVM (97.37% accuracy), Random Forest (96.49%), and a basic neural network (96.49%).
Uncertainty was higher for wrong predictions (1.7× average) and absent data (2.5× higher). The top 10% uncertain cases had 91.67% accuracy, showing good flagging of tricky cases.
Multi-view fusion improved accuracy by 1.75% over single-view models, confirming its value. The model handled imbalances well, with 97.62% sensitivity for malignant cases.
4. Discussion
MUVED-BC stands out by merging multi-view learning with evidential uncertainty, hitting top accuracy while adding safety via uncertainty flags. Unlike traditional models that overcommit on errors, our approach spots uncertain cases for doctor review, boosting trust in AI.
The 99.12% accuracy tops recent WBCD studies, with perfect precision meaning no false alarms for malignant cases. Uncertainty tracking aligns with medical needs, where high uncertainty prompts further tests. Multi-view use leverages WBCD’s structure for better feature capture.
This lightweight model suits low-resource settings. Clinically, it enables AI-human teamwork: AI handles clear cases, flags others for experts, saving time while ensuring safety.
Limitations: Single-dataset testing and retrospective data. Future: Test across hospitals, add real trials, mix with images, and apply to other cancers.
5. Conclusion
MUVED-BC advances safe medical AI by pairing top diagnostic accuracy with reliable uncertainty measures. It flags uncertain cases effectively while acing clear ones, building a base for trusted clinical use. Future focus: Real-world testing and workflow integration to transform cancer detection through smart AI-human partnerships.
Acknowledgments
We used public data from the UCI Machine Learning Repository.
References
[1] Siegel, R. L., Miller, K. D., & Jemal, A. (2020). Cancer statistics, 2020. CA: A Cancer Journal for Clinicians, 70(1), 7–30.
[2] Wolberg, W. H., Street, W. N., & Mangasarian, O. L. (1994). Machine learning techniques to diagnose breast cancer from fine-needle aspirates. Cancer Letters, 77(2–3), 163–171.
[3] Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. International Conference on Machine Learning.
[4] Lakshminarayanan, B., Pritzel, A., & Blundell, C. (2017). Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in Neural Information Processing Systems.
[5] Kompa, B., Snoek, J., & Beam, A. L. (2021). Second opinion needed: communicating uncertainty in medical machine learning. NPJ Digital Medicine, 4(1), 1–6.
[6] Johnson, J. M., & Khoshgoftaar, T. M. (2019). Survey on deep learning with class imbalance. Journal of Big Data, 6(1), 1–54.
[7] Sensoy, M., Kaplan, L., & Kandemir, M. (2018). Evidential deep learning to quantify classification uncertainty. Advances in Neural Information Processing Systems.
[8] Zhao, J., Xie, X., Xu, X., & Sun, S. (2017). Multi-view learning overview: Recent progress and new challenges. Information Fusion, 38, 43–54.
[9] Dua, D., & Graff, C. (2019). UCI Machine Learning Repository. University of California, Irvine.
[10] Street, W. N., Wolberg, W. H., & Mangasarian, O. L. (1993). Nuclear feature extraction for breast tumor diagnosis. IS&T/SPIE’s Symposium on Electronic Imaging: Science and Technology.
[11] Topol, E. J. (2019). High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine, 25(1), 44–56.