An interpretable classifier for high-resolution breast cancer screening images utilizing weakly supervised localization (2002.07613v1)

Abstract: Medical images differ from natural images in significantly higher resolutions and smaller regions of interest. Because of these differences, neural network architectures that work well for natural images might not be applicable to medical image analysis. In this work, we extend the globally-aware multiple instance classifier, a framework we proposed to address these unique properties of medical images. This model first uses a low-capacity, yet memory-efficient, network on the whole image to identify the most informative regions. It then applies another higher-capacity network to collect details from chosen regions. Finally, it employs a fusion module that aggregates global and local information to make a final prediction. While existing methods often require lesion segmentation during training, our model is trained with only image-level labels and can generate pixel-level saliency maps indicating possible malignant findings. We apply the model to screening mammography interpretation: predicting the presence or absence of benign and malignant lesions. On the NYU Breast Cancer Screening Dataset, consisting of more than one million images, our model achieves an AUC of 0.93 in classifying breasts with malignant findings, outperforming ResNet-34 and Faster R-CNN. Compared to ResNet-34, our model is 4.1x faster for inference while using 78.4% less GPU memory. Furthermore, we demonstrate, in a reader study, that our model surpasses radiologist-level AUC by a margin of 0.11. The proposed model is available online: https://github.com/nyukat/GMIC.

An Interpretable Classifier for High-Resolution Breast Cancer Screening Images Utilizing Weakly Supervised Localization

The presented paper explores the usage of neural networks in medical image analysis, specifically for the task of breast cancer screening via high-resolution mammograms. The key challenge addressed is the inherent difference between medical images and natural images, particularly regarding resolution and the size of regions of interest (ROIs).

Model Architecture

The paper extends on the globally-aware multiple instance classifier (GMIC), proposing a novel model that combines a low-capacity global module with a high-capacity local module. This model first processes the entire image to generate saliency maps, highlighting potential regions of interest. Subsequently, a set of patches from these regions are processed with higher detail, and the information is finally aggregated using a fusion module to produce the classification output.

Key Components:

1. Global Module: A low-capacity ResNet-22 network that produces saliency maps after applying it to the entire image.
2. Patch Extraction: ROI patches are selected based on the saliency maps using a designed retrieve_roi algorithm.
3. Local Module: A higher-capacity network (ResNet-18/34/50) processes these patches to capture refined details.
4. Fusion Module: Combines the high-level and detailed features from the two preceding modules to make the final prediction.

Performance Evaluation

The model was evaluated on the NYU Breast Cancer Screening Dataset, which comprises over a million images. Several key metrics were assessed:

1. Classification Accuracy: The model achieved an impressive AUC of 0.93 in identifying malignant findings, outperforming baseline models such as ResNet-34 and Faster R-CNN.
2. Efficiency: Compared to ResNet-34, the model showcased a significant reduction in GPU memory (78.4%) and was faster during inference (4.1x speedup).
3. Reader Study: Compared against human radiologists, the model surpassed the average performance of radiologists by 0.11 in terms of AUC, indicating substantial potential for real-world application.

Key Findings and Implications

1. Fusion Module: The fusion module, which combines global and local features, provided a significant boost in classification performance over models that used only global or local predictions.
2. Patch-wise Attention: Incorporating patch-wise attention further fine-tuned the model’s ability to focus on the most relevant details in the high-resolution images.
3. Aggregation Function: The paper highlighted that top t% pooling for aggregating saliency maps achieved a balance between capturing relevant information and preventing noise, outperforming GAP and GMP.

Practical and Theoretical Implications

From a practical perspective, this model demonstrates significant promise for enhancing diagnostic accuracy in breast cancer screening while reducing the computational burden, potentially leading to significant clinical improvements. The approach of weakly supervised localization ensures that detailed annotations, which are often expensive and time-consuming to collect, are not necessary—broadening the applicability of this model to other domains with similar constraints.

The paper also showcases the robustness of the model through its detailed ablation studies, illustrating the importance of each component and its impact on the overall performance. The model's ability to outperform existing state-of-the-art methods not only in AUC but also in computational efficiency marks a significant step forward.

Future Directions

Future developments could explore dynamic learning schedules for training both the global and local modules more effectively. Additionally, extending this methodology to other imaging modalities like ultrasound or MRI could further verify its versatility. Incorporating learnable pooling thresholds and exploring more sophisticated ROI selection algorithms could refine the model further.

Conclusion

The advancements introduced in this paper offer valuable insights into the specific challenges of high-resolution medical image analysis and propose a thoughtful neural network architecture that addresses these challenges effectively. The GMIC model stands as a promising tool for improving breast cancer screening and demonstrates the potential for broader application across various high-resolution medical imaging tasks.