Method and system for assessing bone age using deep neural network
The present invention relates to a method and system for assessing bone age using deep neural network, more specifically, in which regions of interest (ROIs) even for rotated objects can be more precisely and accurately extracted from an image by a rotated object detection technique used in region proposal networks. Thereby bones with different angles in the image can be detected with excellent speed and accuracy.
1. A method for bone age assessment, comprising:
receiving, by a bone age assessment system, an image to assess a bone age;
generating, by the bone age assessment system, feature maps for the received image through convolutional neural networks;
generating, by the bone age assessment system, at least one of proposed regions of interest (ROIs) using the generated feature maps through region proposal networks;
performing, by the bone age assessment system, at least one of ROI pooling in an ROI pooling layer and ROI aligning in an ROI aligning layer using the proposed ROIs and the generated feature maps;
generating, by the bone age assessment system, feature vectors;
performing, by the bone age assessment system, max pooling for the feature vectors using fully connected layers; and
classifying, by the bone age assessment system, bone maturity levels and positions of the proposed ROIs,
wherein the proposed ROIs comprise a plurality of rotated ROIs, which are rotated by a plurality of different angles,
wherein the proposed ROIs are extracted via a two-step ROIs extraction method, in which bounding ROIs (bROIs) are extracted at first for global features from the feature maps, and then the proposed ROIs included in each bROI are extracted,
wherein the ROI aligning is performed for the bROIs whose sizes are increased,
wherein the ROI aligning is performed by binary interpolation, which is calculated for subcells of each of the proposed ROIs by less than a pixel unit at an actual pixel position of the proposed ROIs, and then max pooling is performed for the binary interpolated subcells, and
wherein a value of each subcell in each proposed ROI in the binary interpolation is calculated by weighed sum of pixels in columns surrounding the subcell.
2. The method of claim 1 , wherein sizes of the bROIs including two or more proposed ROIs are increased vertically to real object directions by constant ratios.
3. A bone age assessment system comprising at least one processor configured to:
receive an image to assess a bone age;
generate feature maps for the received image from the image through convolutional neural networks;
generate at least one of proposed regions of interest (ROIs) using the generated feature maps through region proposal networks;
perform at least one of ROI pooling in an ROI pooling layer and ROI aligning in an ROI aligning layer using the proposed ROIs and the generated feature maps;
generate feature vectors;
perform max pooling for the feature vectors using fully connected layers; and
classify bone maturity levels and positions of the proposed ROIs,
wherein the proposed ROIs comprise a plurality of rotated ROIs, which are rotated by a plurality of different angles,
wherein the proposed ROIs are extracted via a two-step ROIs extraction method, in which bounding ROIs (bROIs) are extracted at first for global features from the feature maps, and then the proposed ROIs included in each bROI are extracted,
wherein ROI aligning is performed for the bROIs whose sizes are increased,
wherein the ROI aligning is performed by binary interpolation, which is calculated for subcells of each of the proposed ROIs by less than a pixel unit at an actual pixel position of the proposed ROIs, and then max pooling is performed for the binary interpolated subcells, and
wherein a value of each subcell in each proposed ROI in the binary interpolation is calculated by weighed sum of pixels in columns surrounding the subcell.
4. The system of claim 3 , wherein sizes of the bROIs including two or more proposed ROIs are increased vertically to real object direction by constant ratios.