Deep Multi-instance Networks with Sparse Label Assignment for Whole Mammogram Classification (1612.05968v1)

Abstract: Mammogram classification is directly related to computer-aided diagnosis of breast cancer. Traditional methods requires great effort to annotate the training data by costly manual labeling and specialized computational models to detect these annotations during test. Inspired by the success of using deep convolutional features for natural image analysis and multi-instance learning for labeling a set of instances/patches, we propose end-to-end trained deep multi-instance networks for mass classification based on whole mammogram without the aforementioned costly need to annotate the training data. We explore three different schemes to construct deep multi-instance networks for whole mammogram classification. Experimental results on the INbreast dataset demonstrate the robustness of proposed deep networks compared to previous work using segmentation and detection annotations in the training.

• The paper proposes a deep learning framework that leverages sparse multi-instance learning to reduce reliance on manual annotation for whole mammogram classification.
• It explores three MIL schemes: max pooling, label assignment, and sparse MIL which applies a sparsity constraint to better capture malignant features.
• Experimental results on the INbreast dataset indicate that the sparse MIL approach outperforms traditional methods by achieving higher AUC while minimizing annotation costs.

Overview of Deep Multi-instance Networks with Sparse Label Assignment for Whole Mammogram Classification

The paper "Deep Multi-instance Networks with Sparse Label Assignment for Whole Mammogram Classification" by Wentao Zhu et al. introduces a methodology aimed at enhancing computer-aided breast cancer diagnosis through a novel deep learning approach applied to whole mammogram classification. The proposed method significantly reduces the reliance on manually annotated training data, which is commonly required by traditional mammogram classification methods.

The authors argue that traditional mammogram classification methods rely heavily on the availability of meticulously annotated datasets, which is both time-consuming and costly. Leveraging deep convolutional features and multi-instance learning (MIL), the authors propose an end-to-end trained deep network that obviates the need for manual labeling. MIL treats each patch within a mammogram as an instance, and the complete mammogram is considered a 'bag' of instances. This approach promises robust whole mammogram classification without elaborate annotation requirements.

Approach and Methodology

Three schemes for constructing deep MIL networks are explored by the authors:

1. Max Pooling-based Multi-instance Learning: This strategy employs a conventional MIL assumption that the presence of at least one positive instance indicates a positive label for the bag. Here, the classification outcome is determined by the instance with the maximum malignant probability.
2. Label Assignment-based Multi-instance Learning: This approach extends the max pooling method by considering a fixed number of top-ranking instances within the bag as representative of the whole mammogram's label, transforming MIL into a label assignment problem.
3. Sparse Multi-instance Learning: By incorporating the prior knowledge that malignant masses are sparse within a whole mammogram, a sparsity constraint is applied to the MIL. This framework balances between the max pooling method's slack assumption and the label assignment method's hard constraints, showing efficacy in whole mammogram classification.

Experimental Results

The methods were evaluated on the INbreast dataset, comprising 410 mammograms with varying results against competing methods. The experiments revealed that the sparse multi-instance network achieved superior accuracy compared to baselines, without the need for explicit detection or segmentation annotations. The sparse and label assignment-based MIL networks, in particular, demonstrated higher Area Under Curve (AUC) values than the traditional max pooling-based approach, showcasing their robustness and ability to utilize the full power of the dataset.

Implications and Future Directions

The findings of this research have both theoretical and practical implications. Theoretically, the transition from a conventional multi-instance approach to label assignment and sparse methodologies enhances our understanding of effective data utilization in deep networks. Practically, this suggests a pathway to developing more cost-effective breast cancer diagnostic systems that require less manual intervention.

Furthermore, future research could extend this model to incorporate multi-scale modeling for enhanced mammogram analysis, or develop adaptive methods for selecting the parameter $k$ in label assignment-based learning. This approach also holds promise for other domains, such as bio-image analysis, where domain expertise is in scarce supply or relevant regions are sparse or small relative to the entire dataset.

In conclusion, the paper presents a notable contribution to the field of medical imaging, specifically in the application of AI for mammogram analysis, suggesting promising avenues for future research efforts in the domain of AI-driven medical diagnostics.