Use Case: AI Breakthrough in Breast Pathology
AI-Powered Discrimination of Breast Fibroepithelial Lesions on Core Biopsies
Use Case: Breast fibroepithelial lesions (FELs), encompassing benign fibroadenomas (FAs) and rarer phyllodes tumors (PTs), pose a significant diagnostic challenge due to their overlapping features on core biopsies. Accurate differentiation is crucial for clinical management, as FAs typically require imaging surveillance or simple enucleation, while PTs necessitate surgical removal with clear margins due to recurrence risks. Conventional core needle biopsies (CNBs) can be hampered by under-sampling, tumor heterogeneity, and inter-observer variability, leading to potential misdiagnoses. This use case highlights the development and evaluation of an artificial intelligence (AI) model designed to enhance diagnostic discrimination between FA and PT on breast core biopsies, aiming for a more objective and rapid tool.
Exact AI Techniques Used: The AI model employs a two-stage architecture combining a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN).
Data Preparation: Whole-slide images (WSIs) of 187 FA and 100 PT core biopsies were utilized. From these, smaller, non-overlapping image patches (224 × 224 pixels) were generated and meticulously filtered to exclude artifacts like ink stains, blur, background, slide edges, folded tissue, adipocytes, and excess blood. Stain variations were normalized. Lesional regions were annotated by pathologists in the training and validation sets to guide the model, but not in the testing set, reflecting real-world clinical application.
CNN Component (Feature Extraction): A ResNet-50 architecture was employed, initialized with weights from the ImageNet database and further trained on augmented training data. This CNN extracted discriminative features from individual image patches. Activation values from the global average pooling layer served as feature representations for the subsequent RNN.
RNN Component (Whole-Slide Classification): The RNN component uses two Long Short-Term Memory (LSTM) layers to aggregate the patch-level features produced by the CNN. Patches were arranged in a row-wise sequence, allowing the RNN to potentially learn spatial relationships and produce an overall classification for the entire WSI.
Training & Performance: Both CNN and RNN models were trained sequentially using a weighted cross-entropy loss function to account for class imbalance. The model achieved an impressive 87.5% overall slide-level accuracy on the unannotated testing subset, with 80% accuracy for FA slides and 95% accuracy for PT slides. Its AUC (Area Under the Receiver Operating Characteristic curve) was 0.875.
Limitations: Despite its promise, the study acknowledges several limitations:
The model currently uses a single magnification factor, potentially missing critical diagnostic features that are better observed at different magnifications.
The RNN architecture models unidirectional sequences of patch-level features along the horizontal axis, which may not fully capture the complex spatial structure of features within a WSI.
The study was limited by a relatively smaller number of PT slides, impacting the FA to PT ratio, though class weights were applied to manage this imbalance.
The use of core biopsy data only (as opposed to larger resection tissue) presents a challenge, as core biopsies offer less information and may not show key diagnostic features like leafy fronds as clearly.
Possible Advancements in Accuracy: Future studies could enhance the model's accuracy by:
Implementing multiple CNN models with inputs at different magnification factors to capture a wider range of discriminative features.
Investigating bidirectional and/or two-dimensional LSTM models to more effectively model the spatial relationships of features across the WSI.
Increasing the size and diversity of the dataset, particularly by acquiring more PT cases, to improve the model's generalization ability.
Performing annotations on more cases to provide richer training data.
Evaluating the model on samples from other institutions/laboratories to assess its broader applicability.
Future Implications: This study affirms the potential of AI to significantly impact diagnostic pathology, especially for breast cancer. For breast cancer patients, this translates to:
Earlier and more accurate pre-operative diagnosis of FELs, potentially reducing the need for surgical excision for FAs. This could lead to significant cost savings, reduced patient anxiety, and fewer unnecessary invasive procedures.
A more objective and rapid diagnostic tool for pathologists, addressing issues of reproducibility and inter-observer variability common in manual assessments.
Further integration of computational and digital pathology into routine diagnostics, accelerating the adoption of AI in clinical settings.
The model's evaluation on unannotated WSIs makes its performance more reflective of real-world clinical use-cases, positioning it for future refinement and application in routine practice.
Full paper here: https://www.laboratoryinvestigation.org/article/S0023-6837(22)00063-0/fulltext
