Heat Map Pre-NMS Filtering for Digital Breast Tomography
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection techniques, such as Non-Maximum Suppression (NMS), face challenges in effectively selecting bounding boxes from multiple detectors, particularly in Digital Breast Tomography, where high-score boxes from few detectors may not overlap with others, leading to suboptimal detection results.
Innovation Solution
The proposed solution involves generating a heatmap where each pixel represents the number of overlapping bounding boxes and their scores, filtering boxes using this heatmap with a threshold proportional to the maximum value, and performing Non-Maximum Suppression on the filtered candidates to output the most accurate bounding box.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If Non-Maximum Suppression is applied to select bounding boxes from multiple detectors, then detection speed is improved, but detection accuracy deteriorates because high-score boxes from few detectors may be incorrectly rejected due to lack of overlap with other boxes
Solution Approach 1:
The patent transitions from traditional NMS that operates in box-space (comparing bounding boxes directly) to a heatmap-based approach that projects detections into pixel-space. Each detector vote is accumulated in a heatmap at the corresponding image location, creating a density map where detection confidence is represented by pixel intensity values rather than box scores. This dimensional transformation allows the system to capture spatial concentration of detections without requiring box overlap, thereby maintaining detection accuracy while preserving computational efficiency.
2Device complexity
If traditional Non-Maximum Suppression is used to choose one bounding box per object, then computational complexity is reduced, but false rejections increase when detectors disagree on box locations
Solution Approach 1:
The patent merges multiple detector votes by accumulating them in a shared heatmap structure. Instead of processing each detector independently and then suppressing overlaps, all detector outputs are combined into a single heatmap where each pixel value represents the cumulative confidence from all detectors voting for that location. This merging approach allows the system to leverage consensus from multiple detectors while maintaining linear computational complexity, as the heatmap accumulation is a simple additive operation that scales linearly with the number of detectors.
Data Source
AI summary
Embodiments may include novel techniques to improve detection of objects in images, for example, in Digital Breast Tomography and that are applicable to ensembles of detectors. For example, a method may comprise generating a plurality of candidate bounding boxes for each of a plurality of image slices of imaged tissue, each generated candidate bounding box having a probability score, collecting at least some of the generated candidate bounding boxes for each slice, generating a heat map of the filtered candidate bounding boxes and filtering the candidate bounding boxes in the heat map based on a first threshold of values in the heat map, performing Non-Maximum Suppression on the heat map filtered candidate bounding boxes, and outputting at least one bounding box based on the probability score.