TL;DR
This paper introduces a combined approach for classifying breast masses and normal tissues in digital tomosynthesis mammography (DTM) by integrating 3D reconstructed data with 2D projection views. This novel method significantly improves classification accuracy, offering a promising tool for reducing false positives in automated breast cancer detection systems.
What Problem Does This Paper Address?
Mammography, while a crucial tool for breast cancer screening, faces limitations due to overlapping dense fibroglandular tissue, which can obscure tumors and lead to missed diagnoses. Digital tomosynthesis mammography (DTM) offers a solution by providing quasi-3D structural information, potentially improving detection and reducing the impact of tissue overlap. However, effectively distinguishing between actual breast masses and normal tissue in DTM data remains a challenge for computer-aided detection (CADe) systems. This paper aims to develop and evaluate an effective classifier to address this challenge, specifically focusing on reducing false positives in DTM-based mass detection.
Methods
The researchers developed a CADe system with five processing steps. For classification, they evaluated three distinct approaches:
- 3D Approach: This method utilizes morphological and run-length statistics texture features extracted directly from the 3D reconstructed DTM slices. A Linear Discriminant Analysis (LDA) classifier is then trained using these features.
- 2D Approach: This approach preprocesses the raw 2D projection views (PVs) using a Laplacian pyramid multi-resolution enhancement. Mass candidates are then forward-projected onto these enhanced PVs to identify corresponding regions of interest (ROIs). Spatial gray-level dependence (SGLD) texture features are extracted from these ROIs and averaged across all PVs. An LDA classifier is trained on these averaged features.
- Combined Approach: This method leverages the strengths of both the 3D and 2D approaches. It merges the LDA scores (representing the likelihood of a candidate being a mass) from the 3D and 2D classifiers, typically through simple averaging, to generate a final mass likelihood score.
A leave-one-case-out resampling method was used for training and testing the classifiers, ensuring robust evaluation. Receiver Operating Characteristic (ROC) analysis was employed to compare the performance of the three approaches.
Datasets
The study utilized a dataset of 100 DTM cases collected from a GE first-generation prototype DTM system at the Massachusetts General Hospital. This dataset included 69 malignant and 31 benign masses, along with some large-area architectural distortions. The DTMs were reconstructed using a simultaneous algebraic reconstruction technique (SART). The paper does not introduce a new benchmark dataset but uses a real-world clinical dataset.
Key Insights and Ideas
- Synergy of 2D and 3D: The most significant insight is that combining information from both the 3D reconstructed DTM slices and the 2D projection views leads to a substantial improvement in classification performance. This suggests that neither modality alone captures all the necessary discriminative information.
- Forward Projection as a Bridge: The novel use of forward projection to map mass candidates from the 3D volume onto the 2D projection views is a key methodological contribution. This allows for the extraction of relevant 2D texture features that complement the 3D information.
- DTM's Potential for FP Reduction: The results strongly indicate that DTM, when analyzed with sophisticated classification techniques like the combined approach, can significantly improve the ability to differentiate masses from normal tissue, which is crucial for reducing false positives in CADe systems.
Key Results
The combined approach demonstrated superior performance compared to the individual 3D and 2D methods:
- Area Under the ROC Curve (A<sub>z</sub>):
- 3D approach: 0.87 ± 0.02
- 2D approach: 0.86 ± 0.02
- Combined approach: 0.91 ± 0.02
- Statistical Significance: The difference in A<sub>z</sub> values between the 3D and 2D approaches was not statistically significant (p=0.50). However, the combined approach's performance was significantly better than either the 3D (p=0.04) or the 2D (p=0.03) approach alone.
Limitations and Future Work
- Limitations:
- The study used a relatively small dataset of 100 DTM cases.
- The performance of the 3D and 2D approaches individually was not significantly different, suggesting that further refinement of feature extraction or classification within each modality might be beneficial.
- The "combined approach" primarily relied on averaging LDA scores, which might not be the optimal fusion strategy.
- Future Work: The authors suggest that the combined classifier will be useful for false-positive reduction in computerized mass detection in DTM. Further research could explore more sophisticated fusion techniques beyond simple averaging and investigate the performance on larger, more diverse datasets. Additionally, exploring other feature extraction methods for both 2D and 3D data could yield further improvements.
Summary updated