Improving Computer-aided Detection using Convolutional Neural Networks and Random View Aggregation (1505.03046v2)

Abstract: Automated computer-aided detection (CADe) in medical imaging has been an important tool in clinical practice and research. State-of-the-art methods often show high sensitivities but at the cost of high false-positives (FP) per patient rates. We design a two-tiered coarse-to-fine cascade framework that first operates a candidate generation system at sensitivities of $\sim$100% but at high FP levels. By leveraging existing CAD systems, coordinates of regions or volumes of interest (ROI or VOI) for lesion candidates are generated in this step and function as input for a second tier, which is our focus in this study. In this second stage, we generate $N$ 2D (two-dimensional) or 2.5D views via sampling through scale transformations, random translations and rotations with respect to each ROI's centroid coordinates. These random views are used to train deep convolutional neural network (ConvNet) classifiers. In testing, the trained ConvNets are employed to assign class (e.g., lesion, pathology) probabilities for a new set of $N$ random views that are then averaged at each ROI to compute a final per-candidate classification probability. This second tier behaves as a highly selective process to reject difficult false positives while preserving high sensitivities. The methods are evaluated on three different data sets with different numbers of patients: 59 patients for sclerotic metastases detection, 176 patients for lymph node detection, and 1,186 patients for colonic polyp detection. Experimental results show the ability of ConvNets to generalize well to different medical imaging CADe applications and scale elegantly to various data sets. Our proposed methods improve CADe performance markedly in all cases. CADe sensitivities improved from 57% to 70%, from 43% to 77% and from 58% to 75% at 3 FPs per patient for sclerotic metastases, lymph nodes and colonic polyps, respectively.

• The paper introduces a two-tier method combining candidate generation with ConvNet classification to enhance CADe sensitivity.
• It employs random view augmentation using 2D/2.5D transformations to effectively lower false-positive rates.
• Experiments on sclerotic metastases, lymph nodes, and colonic polyps demonstrate significant improvements in detection accuracy.

Improving Computer-Aided Detection Using Convolutional Neural Networks and Random View Aggregation

The paper focuses on advancements in automated computer-aided detection (CADe) using Convolutional Neural Networks (ConvNets) combined with random view aggregation for medical imaging. The core contribution is a two-tiered framework designed to enhance CADe sensitivity while maintaining manageable false-positive rates, a typical challenge in existing systems.

Proposed Methodology

The proposed approach consists of a two-tier cascade system:

1. Candidate Generation: Initial identification of potential Regions of Interest (ROIs) or Volumes of Interest (VOIs) with high sensitivity but high false positives.
2. ConvNet Classification: The focal point of the research, employing ConvNets trained on randomly sampled 2D or 2.5D views derived from the ROIs/VOIs generated in the first tier. This involves transformations such as scale, translation, and rotation, enhancing robustness and classification accuracy.

Evaluation and Experimental Results

The methodology was validated using three datasets encompassing sclerotic metastases, lymph nodes, and colonic polyps, with the number of patients being 59, 176, and 1,186 respectively. Noteworthy numerical results demonstrate significant performance improvements in CADe systems. Sensitivities increased from:

• 57% to 70% for sclerotic metastases
• 43% to 77% for lymph nodes
• 58% to 75% for colonic polyps at 3 false positives per patient

These improvements are driven by random view augmentation and aggregation, optimizing ConvNet performance even in data-scarce scenarios.

Significance and Theoretical Implications

This research has multiple implications:

• Enhanced CADe Sensitivity: By integrating ConvNets, the method shows marked improvements in accurately detecting pathologies with fewer false positives.
• Versatility and Scalability: The universal 2.5D image decomposition allows adaptation to various medical imaging applications beyond those tested.
• Efficient Training: Random view aggregation presents a data augmentation strategy that mitigates overfitting, crucial in medical datasets with limited samples.

Future Directions

Future research could explore:

• Deeper ConvNet Architectures: Extending the depth and complexity of ConvNets could potentially yield further gains, albeit with increased computational demands.
• 3D ConvNet Exploration: While challenging due to computational constraints, direct applications of 3D ConvNets might offer insights into volumetric imaging analysis.
• Transfer Learning: Leveraging pre-trained models could simplify initialization, potentially enhancing performance with less training data.

Conclusion

The integration of ConvNets with random view aggregation presents a robust methodology for improving CADe systems across multiple medical imaging applications. This paper illustrates the capacity of deep learning to deliver high sensitivity and specificity, marking a significant step forward in clinical AI applications. Employing comprehensive evaluation, the framework offers a promising avenue for further enhancing diagnostic precision in medical practices.