TL;DR
This paper introduces a novel, self-interpretable method for classifying invasive breast cancer in histopathology images. By learning patterns from normal tissue and identifying deviations, it generates a probability map highlighting suspicious regions, outperforming existing methods with less reliance on extensive annotations.
What Problem Does This Paper Address?
Accurate diagnosis of breast cancer from histopathology images is crucial but challenging due to the complex and varied nature of cancerous cell growth and benign tissue abnormalities. Existing deep learning approaches often require extensive, pixel-level annotations, which are time-consuming and expensive to obtain. Furthermore, many current methods lack "self-interpretability," meaning they can't explain why they make a particular diagnosis, hindering trust and clinical adoption. This paper aims to bridge this gap by developing a method that requires only image-level labels (normal vs. malignant) and provides a visual explanation of its diagnostic reasoning.
Methods
The proposed method employs a three-phase approach:
- Contrast Pattern Mining with a Fully Convolutional Autoencoder (FCAE): The core idea is to learn what "normal" breast tissue looks like. An FCAE is trained only on patches from normal histopathology images. This autoencoder is designed to reconstruct these normal patches. The key innovation here is using the Structural Similarity Index Measure (SSIM) as the loss function instead of the more common Mean Squared Error (MSE). SSIM helps preserve structural information, leading to sharper reconstructions and better capture of subtle patterns. The reconstruction residues (the difference between the input and reconstructed patch) are then analyzed. Normal patches will have low residues, while patches deviating from the learned normal patterns will have high residues.
- Patch Labeling with One-Class Support Vector Machine (OC-SVM): The reconstruction residues from the FCAE are then used to generate feature descriptors. An OC-SVM is trained on these features from normal patches to define a "normal" region in the feature space. Patches whose features fall outside this normal region are flagged as potentially abnormal. This is a form of anomaly detection.
- Malignant Tumor Probability Map Generation with a 1-Layer Neural Network (NN): The labels generated by the OC-SVM are then used to train a simple 1-layer neural network. This network learns to output a probability score for each patch, indicating the likelihood of it containing malignant cells. This probability score is then used to generate a heat-map (probability map) for the entire image, visually highlighting areas of suspected abnormality.
Datasets
The study utilizes a public breast cancer biopsy image set from the Israel Institute of Technology. This dataset contains 361 samples, classified into normal tissue (119 images), carcinoma in situ (102 images), and invasive carcinoma (140 images). The images are H&E stained and captured at 40x magnification.
Key Insights and Ideas
- Learning "Normal" to Detect "Abnormal": Instead of directly learning malignant patterns (which are highly variable), the method focuses on learning the robust patterns of normal tissue. Deviations from this learned normality are then indicative of abnormality. This is a powerful unsupervised learning strategy.
- SSIM for Structural Preservation: Using SSIM as the loss function for the autoencoder is crucial. It ensures that the learned representations capture the intricate structural details of normal tissue, which is vital for distinguishing subtle differences in histopathology.
- Self-Interpretability through Probability Maps: The generated probability maps provide a visual explanation for the diagnosis. This allows pathologists to see where the algorithm is focusing its attention, fostering trust and enabling them to verify the AI's reasoning.
- Weak Supervision is Sufficient: The method demonstrates that image-level labels (normal/malignant) are sufficient for training, significantly reducing the annotation burden compared to methods requiring pixel-level annotations.
Key Results
The proposed method achieved superior classification performance compared to two state-of-the-art patch-based deep learning methods (Spanhol et al. and Araujo et al.) on the public dataset. Specifically, it reported higher Accuracy (0.760 vs. 0.700/0.710), F1-score (0.777 vs. 0.766/0.763), Likelihood Ratio+ (2.645 vs. 1.540/1.554), and Diagnostic Odds Ratio (12.876 vs. 8.563/8.132). The probability maps generated by the proposed method were also qualitatively shown to be more localized to actual abnormality regions compared to the other methods, which tended to highlight background areas.
Limitations and Future Work
The authors acknowledge that the primary limitation of their study is the size of the experimental image set and the absence of external validation. They also note that while their method offers self-interpretability, the "black-box" nature of deep learning still presents challenges. Future work could involve investigating more efficient pattern mining methods to handle the heterogeneity of histological abnormalities and exploring techniques to further interpret the internal reasoning of deep diagnosis systems, potentially by building upon methods like CAM and Grad-CAM.
Summary updated