Applications of Artificial Intelligence in Breast Pathology

Introduction to AI in Breast Pathology

• Context: Increasing use of whole slide imaging combined with digital workflows allows for incorporating AI into pathology, particularly in breast pathology.

• Breast pathologists deal with:

  • High workloads

  • Complex diagnoses

  • Repetitive tasks

  • Semi-quantitative evaluations of biomarkers

• AI advancements have provided promising solutions to these challenges.

• Objective: Review current and potential AI applications in breast pathology focusing on diagnoses, grading breast carcinomas, detecting lymph node metastasis, quantifying biomarkers, predicting prognosis and treatment responses, and potential molecular changes.

The Role of Pathology in Disease Diagnosis

• Pathology serves as the gold standard for diagnosing diseases, especially malignant tumors.

• Advances in personalized treatment necessitate high standards in pathologic diagnosis.

• Pathologic diagnosis includes:

  • Morphologic identification of tumor cells

  • Evaluation of mitotic counts

  • Determining lymph node metastasis

  • Interpretation of biomarkers

• Challenges in Pathology: Limitations include:

  • Shortage of trained pathologists

  • Variability in diagnostic skills and interpretation

  • Reliance on ancillary studies.

Whole Slide Imaging and AI Implementation

• Whole slide imaging (WSIs) and digital workflows transform pathology by enabling:

  • Digital sign-out and remote consultations

  • Efficient slide storage for research and education.

• AI applications address workload issues, enhance diagnostic accuracy, and provide insights into treatment responses and prognoses.

• Machine Learning (ML): A subfield of AI where algorithms learn data patterns from large datasets to predict outcomes in new cases.

• Deep Learning (DL): A subset of ML utilizing neural networks for complex data pattern recognition; includes supervised, semi-supervised, unsupervised, and transfer learning techniques.

  • Deep Convolutional Neural Networks (DCNN): Particularly effective for image recognition and analysis.

Current and Potential AI Applications in Breast Pathology

1. AI-based Detection and Classification

• Accurate classification and grading of breast cancers are critical for clinical management.

• Recent research includes:

  • Novel DL models for classifying multiple histopathologic types (e.g., ductal carcinoma, lobular carcinoma).

  • High performance achieved (e.g., accuracy of 93.2%).

  • Models identifying invasive carcinoma with high accuracy through examined data.

• Commercially available AI platforms can screen image blocks to generate heat maps indicating different lesion types.

  • Example: GALEN Breast algorithm: AUC results indicate effective detection of invasive carcinoma with AUC of 0.99.

2. AI-based Grading of Breast Carcinoma

• The Nottingham grading system comprises tubule formation, nuclear pleomorphism, and mitotic activity.

• AI neural networks can improve grading accuracy by:

  • Automating mitotic counts.

  • Identifying nuclear characteristics objectively.

  • Reducing interobserver variability.

• Examples of AI Systems for Grading:

  • MaskMitosis: Outperformed existing mitosis detection approaches with a significant F-score.

  • Algorithms validated across several datasets for accurate detection of key morphological features.

3. Detection of Lymph Node Metastasis

• Accurate assessment of lymph node status is crucial for management and prognosis in breast cancer patients.

• AI can streamline evaluation processes:

  • Challenges noted for detecting micrometastasis and isolated tumor cells.

  • AI algorithms have shown higher detection accuracies compared to traditional methods, as illustrated by several studies and competitions.

• Commercial applications, such as Visiopharm, facilitate comprehensive assessment workflows.

4. Breast Cancer Biomarker Quantification

• Tissue biomarkers serve as indicators for diagnosis and management.

• AI methodologies improve accuracy and reproducibility in assessing biomarkers like ER, PR, HER2, and Ki-67 through:

  • Automated image analysis in correlation with pathologists’ manual scoring.

  • AI algorithms calculate positivity rates, intensity of staining, and can identify features like nuclear morphology.

5. Prediction of Prognosis and Therapy Response

• Identifying patients who significantly benefit from therapies, such as chemotherapy, is essential to minimize toxicity.

• AI algorithms, like IbRiS, utilize histopathologic data to predict outcomes comparable to Oncotype DX with an accuracy of 84%.

• Integration of histopathological data and biomarkers can yield additional insights into patient management.

Limitations of AI in Breast Pathology

• Data Quality: The accuracy of AI systems heavily relies on the quality of training datasets.

• Validation: Thorough validation of AI algorithms is necessary for clinical use. Many noted studies remain experimental.

• Digital Workflow Implementation: The integration of AI needs robust digital pathology infrastructure.

• Pathologists’ Trust: Many AI models operate as 'black boxes', making it critical for pathologists to understand how algorithms derive outcomes to promote acceptance.

Conclusion

• The application of AI is growing within breast pathology, promising to reduce workloads of repetitive and complex tasks.

• AI can provide tools for new assessments including TIL quantification and therapy response predictions, and may complement or replace certain costly molecular tests in routine practice.