High-Resolution Breast Cancer Screening with Multi-View Deep Convolutional Neural Networks (1703.07047v3)

Abstract: Advances in deep learning for natural images have prompted a surge of interest in applying similar techniques to medical images. The majority of the initial attempts focused on replacing the input of a deep convolutional neural network with a medical image, which does not take into consideration the fundamental differences between these two types of images. Specifically, fine details are necessary for detection in medical images, unlike in natural images where coarse structures matter most. This difference makes it inadequate to use the existing network architectures developed for natural images, because they work on heavily downscaled images to reduce the memory requirements. This hides details necessary to make accurate predictions. Additionally, a single exam in medical imaging often comes with a set of views which must be fused in order to reach a correct conclusion. In our work, we propose to use a multi-view deep convolutional neural network that handles a set of high-resolution medical images. We evaluate it on large-scale mammography-based breast cancer screening (BI-RADS prediction) using 886,000 images. We focus on investigating the impact of the training set size and image size on the prediction accuracy. Our results highlight that performance increases with the size of training set, and that the best performance can only be achieved using the original resolution. In the reader study, performed on a random subset of the test set, we confirmed the efficacy of our model, which achieved performance comparable to a committee of radiologists when presented with the same data.

High-Resolution Breast Cancer Screening with Multi-View Deep Convolutional Neural Networks

The paper "High-Resolution Breast Cancer Screening with Multi-View Deep Convolutional Neural Networks" advances the integration of deep learning methodologies in medical imaging, specifically targeting breast cancer screening through mammography. The authors propose a novel approach using a multi-view deep convolutional neural network (MV-DCN) capable of handling high-resolution medical images without the necessity to downscale them, thus preserving critical fine details essential for accurate medical diagnoses.

Contributions of the Study

The primary contribution of this paper lies in the development of the MV-DCN, which processes multiple high-resolution views of mammographic screenings and predicts the Breast Imaging Reporting and Data System (BI-RADS) categories with high accuracy. Importantly, the network is trained on a substantial dataset comprising 886,437 images, a scale that significantly exceeds prior studies in the domain. The paper highlights the critical importance of using large-scale datasets to harness the full potential of deep neural networks in medical imaging tasks.

Numerical Results and Model Performance

A key finding presented in the paper is that the performance of the MV-DCN improves with an increase in both the size of the training dataset and the resolution of input images. Specifically, the paper reports that the use of original image resolution, as opposed to downscaled versions, is pivotal for achieving the best performance due to the necessity of capturing detailed image nuances vital for detecting subtle pathological signs such as microcalcifications or tissue asymmetry.

The performance of the proposed model was evaluated using a reader paper test set, where the model's accuracy was benchmarked against the assessments of a committee of experienced radiologists. The MV-DCN demonstrated a macAUC of 0.733, approaching the performance of the radiologists. Notably, the paper documented that ensemble predictions - combining the outputs of the MV-DCN and the radiologists' committee - further improved classification accuracy.

Theoretical and Practical Implications

The incorporation of high-resolution images in deep learning for medical imaging underscores the necessity to address computational and memory challenges, which the current paper successfully mitigates through novel network architecture designs, such as aggressive convolution and pooling techniques. The paper's approach demonstrates the feasibility of deploying end-to-end trainable models in clinical environments, suggesting potential for these models to not only augment radiologists' analyses but also, with further refinement and trust establishment, operate autonomously in certain screening scenarios.

Additionally, the various outcomes observed, in terms of performance related to confidence scores and inter-radiologist agreement, reflect the complexity and inherent challenges of medical image interpretation. This suggests the future potential of deploying such systems in supportive roles, especially considering the model's ability to accurately capture specific pathological insights visually confirmed through relevance mapping techniques.

Future Directions

For further research and development, the critical aspect of realistic, fine-grained label acquisition reflecting actual cancer progression could optimize model learning, beyond the conventional reliance on BI-RADS categories that often vary in consistency. Moreover, expanding the dataset size and conducting thorough hyperparameter optimization could further refine model performance. In practice, this research paves the way for broader applications across various forms of cancer screening and potentially other medical imaging tasks necessitating fine-detail interpretation.

In conclusion, this paper offers a comprehensive and methodologically sound exploration of high-resolution deep learning applications in breast cancer screening. Through its large-scale deployment and validation, it sets a precedent for future research aimed at achieving unparalleled levels of precision and efficacy in medical diagnosis via advanced computational methods.