An interpretable semi-supervised framework for patch-based classification of breast cancer.

Area of Science:

• Computational pathology
• Machine learning in oncology
• Breast cancer diagnostics

Background:

• Accurate detection of invasive Ductal Carcinoma (IDC) is critical for breast cancer diagnosis but challenging.
• Deep neural networks (DNNs) show promise but require extensive labeled data, which is costly and time-consuming to acquire.
• The need for efficient methods that leverage readily available unlabeled data is significant.

Purpose of the Study:

• To develop a novel semi-supervised learning framework for improved IDC detection.
• To reduce reliance on large amounts of manually labeled training data by utilizing unlabeled data.
• To enhance the interpretability and trustworthiness of the DNN predictions.

Main Methods:

• A five-stage framework including data augmentation, feature selection, co-training data labeling, DNN modeling, and prediction interpretability.
• Utilized digitized histopathology slides from 162 women with IDC.
• Compared the proposed framework against state-of-the-art methods like AlexNet, VGG, DCGAN, and self-learning techniques.

Main Results:

• The proposed DNN approach achieved a balanced accuracy of 0.73 and an F-measure of 0.843, outperforming AlexNet and VGG trained solely on labeled data.
• The semi-supervised framework demonstrated superior performance compared to DCGAN and self-learning techniques.
• Achieved 85.75% accuracy, 0.865 balanced accuracy, and 0.773 F-measure using only 10% labeled data.

Conclusions:

• The novel semi-supervised learning framework effectively detects IDC using minimal labeled data, significantly reducing annotation costs.
• The approach offers a promising solution for breast cancer diagnosis, enhancing accuracy and efficiency.
• The inclusion of interpretability enhances clinical trust and adoption of AI-driven diagnostic tools.