IAIA-BL: A Case-based Interpretable Deep Learning Model for Classification of Mass Lesions in Digital Mammography (2103.12308v1)

Abstract: Interpretability in machine learning models is important in high-stakes decisions, such as whether to order a biopsy based on a mammographic exam. Mammography poses important challenges that are not present in other computer vision tasks: datasets are small, confounding information is present, and it can be difficult even for a radiologist to decide between watchful waiting and biopsy based on a mammogram alone. In this work, we present a framework for interpretable machine learning-based mammography. In addition to predicting whether a lesion is malignant or benign, our work aims to follow the reasoning processes of radiologists in detecting clinically relevant semantic features of each image, such as the characteristics of the mass margins. The framework includes a novel interpretable neural network algorithm that uses case-based reasoning for mammography. Our algorithm can incorporate a combination of data with whole image labelling and data with pixel-wise annotations, leading to better accuracy and interpretability even with a small number of images. Our interpretable models are able to highlight the classification-relevant parts of the image, whereas other methods highlight healthy tissue and confounding information. Our models are decision aids, rather than decision makers, aimed at better overall human-machine collaboration. We do not observe a loss in mass margin classification accuracy over a black box neural network trained on the same data.

• The paper introduces IAIA-BL, an interpretable deep learning model using case-based reasoning to classify mass lesions in digital mammography, addressing limitations of black-box models and small medical datasets.
• A key methodological contribution is the model's ability to effectively integrate mixed-label training data, utilizing both image-level and pixel-wise annotations to enhance accuracy and focus on medically relevant regions.
• IAIA-BL achieved predictive performance comparable to black-box models (AUROC) while demonstrating high alignment with radiologists' reasoning (Cohen's Kappa 0.74) and superior interpretability using activation precision.

Interpretable Deep Learning for Mammography: A Case-Based Approach

In this paper, the authors propose "IAIA-BL," an interpretable deep learning framework designed for classifying mass lesions in digital mammography. This model endeavors to address pivotal challenges in medical imaging, particularly the small size of datasets, the presence of confounding variables, and the inherent difficulty in discerning between malignant and benign lesions. In contrast to opaque black-box models, IAIA-BL emphasizes transparency, allowing practitioners to better understand the AI's decision-making process by emulating a radiologist’s reasoning.

Methodological Contributions

1. Interpretable Neural Networks: The paper introduces a framework that uses case-based reasoning within neural networks to construct interpretable models that replicate the cognitive processes of radiologists. This methodology is distinguished by its use of Prototypical Part Network (ProtoPNet), augmented by a new mechanism that provides explanations for the model's decisions. These explanations help identify relevant areas associated with specific medical features, thereby increasing the model's trustworthiness.
2. Mixed-Label Training Data Utilization: An innovative facet is the model's ability to integrate data with both image-level labels and pixel-wise annotations. This flexibility enhances accuracy and interpretability, even with limited datasets. By utilizing fine annotations from a section of images, the framework focuses the model’s attention on medically significant regions, reducing the possibility of decision-making based on confounding information.
3. Top-k Average Pooling: The integration of a top-k average pooling strategy, instead of the traditional max-pooling used in ProtoPNet, significantly boosts performance. This modification allows the model to average the top k closest matches rather than relying solely on a single closest match, enhancing both robustness and interpretability.

Performance and Interpretability

The model’s predictive capability is notable, achieving an Area Under the Receiver Operating Characteristic curve (AUROC) comparable to leading black-box models, without foregoing interpretability. Notably, the model maintained a high alignment with radiologists' reasoning, demonstrated by a Cohens Kappa of 0.74, which surpasses typical inter-physician agreement levels seen in similar tasks.

IAIA-BL exhibited superior interpretability—quantified using the novel metric of activation precision—highlighting relevant image segments more accurately as compared to baseline methods, which often depend on confounding elements.

Implications and Future Work

The implications of deploying IAIA-BL in clinical settings are manifold. Beyond its capability as a decision aid, this model sets a precedent for transparent AI participation in healthcare, allowing human practitioners to validate AI-based assessments. By illuminating its internal reasoning, IAIA-BL can help mitigate risks associated with the deployment of AI in critical healthcare applications.

Furthermore, future research directions include expanding the model's capability to incorporate additional BI-RADS features and potentially extending its application to digital breast tomosynthesis. Reader studies examining the interpretability and accuracy of IAIA-BL in practical radiological assessments would significantly bolster its clinical credibility.

IAIA-BL represents an evolution in AI-assisted diagnostics by bridging the gap between accurate predictions and interpretability, thus fostering collaboration between AI and radiologists in high-stakes medical decision-making.