Lesion Detection via Feature Map Resampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning methods for detecting and locating lesions in medical images require a large workload for marking accurate box information during model training and are inefficient in updating and learning due to the need for extensive manual annotation of sample images.
Innovation Solution
A method utilizing a pre-trained deep learning model that inputs a target medical image to generate a confidence sequence for each preset lesion type, extracts feature maps, calculates fusion feature maps, and resamples them to locate lesions by determining the maximum connected area and creating a mark border, thereby reducing the marking workload and enabling rapid model updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional deep learning methods are used for detecting and locating lesions in medical images, then accurate classification and recognition of lesion types can be achieved, but the marking workload becomes extremely large and model updating becomes inefficient
Solution Approach 1:
The patent segments the lesion detection task into two independent parts: (1) a pre-trained deep learning model that performs classification and provides feature maps without requiring location annotations, and (2) a post-processing module that automatically generates location boxes by thresholding and processing the feature maps. This segmentation allows the system to achieve accurate lesion detection while eliminating the need for manual location marking during model training, thereby improving productivity.
Solution Approach 2:
The patent introduces feature maps as an intermediary between the pre-trained deep learning model and the final location detection. The feature maps serve as a mediator that contains both classification information and spatial information, allowing the system to derive location data without requiring direct annotation of bounding boxes during training. This intermediary enables efficient model updating while maintaining detection accuracy.
2Loss of information
If manual marking of accurate box information is performed during model training, then location information of lesions can be obtained, but the marking workload becomes huge and time-consuming
Solution Approach 1:
The patent implements self-service by enabling the system to automatically generate location information from feature maps without requiring manual annotation. The post-processing module automatically thresholds the feature maps, identifies connected components, and generates bounding boxes, thereby obtaining complete location information while eliminating the time-consuming manual marking process.
Solution Approach 2:
The patent performs preliminary action by pre-training the deep learning model to extract meaningful feature maps that inherently contain spatial information about lesions. These pre-extracted feature maps serve as a foundation that enables automatic location detection in subsequent steps, eliminating the need for manual marking during both training and inference phases.
3Measurement precision
If extensive manual annotation of sample images is performed, then the model can be trained with accurate location data, but rapid updating and learning becomes difficult
Solution Approach 1:
The patent applies parameter changes by transforming the training requirements from manual bounding box annotations to automated feature map processing. The system trains on classification labels only, and automatically generates location parameters through post-processing of feature maps. This parameter change enables rapid model updating and adaptability to new data without requiring re-annotation of existing samples.
Solution Approach 2:
The patent introduces dynamics by making the location generation process adaptive and flexible. The post-processing module dynamically adjusts threshold values and processing parameters based on the input feature maps, allowing the system to adapt to different lesion types and imaging conditions. This dynamic approach enables rapid model updating and versatility without requiring manual re-marking of samples.
Data Source
AI summary
A method for detecting and locating a lesion in a medical image is provided. A target medical image of a lesion is obtained and input into a deep learning model to obtain a target sequence. A first feature map output from the last convolution layer in the deep learning model is extracted. A weight value of each network unit corresponding to each preset lesion type in a fully connected layer is extracted. For each preset lesion type, a fusion feature map is calculated according to the first feature map and the corresponding weight value and resampled to the size of the target medical image to generate a generic activation map. The maximum connected area in each generic activation map is determined, and a mark border surrounding the maximum connected area is created. A mark border corresponding to each preset lesion type is added to the target medical image.