Clinical reading of lateral cervical spine radiographs often requires manual adjustment of contrast to highlight the characteristics of the regions of interest. Automatic enhancement schemes can replace manual labor but often introduce issues including over-enhancement, under-enhancement, increased background noise, and an inability to separate Region of Interest (ROI) from the complex-structured background. This study proposes an identification and enhancement scheme for regions of interest in lateral cervical spine radiographs. First, a target detector for the regions of C1-C7 vertebrae was trained on 1,000 images. Utilizing the explanatory power of the HiResCAM on convolutional neural networks, Class Activation Mapping of shallow features was employed to automatically delineate the vertebrae regions. Subsequently, local contrast enhancement was conducted to improve the readability of the vertebral areas while avoiding the amplification of background noise. Pixel-level alpha blending was applied at the edges of the ROI to eliminate the boundaries between regions. Experimental results on real medical images demonstrate that this method significantly enhances the readability of the vertebral areas in lateral cervical spine images without damaging peripheral information, surpassing previous approaches with its human-like focus on the local ROIs.
Class Activation Mapping-guided Delineation of ROI in Medical Images for Automatic Local Contrast Enhancement in Lateral Cervical Spine Radiographs
Zhang M.;
2024-01-01
Abstract
Clinical reading of lateral cervical spine radiographs often requires manual adjustment of contrast to highlight the characteristics of the regions of interest. Automatic enhancement schemes can replace manual labor but often introduce issues including over-enhancement, under-enhancement, increased background noise, and an inability to separate Region of Interest (ROI) from the complex-structured background. This study proposes an identification and enhancement scheme for regions of interest in lateral cervical spine radiographs. First, a target detector for the regions of C1-C7 vertebrae was trained on 1,000 images. Utilizing the explanatory power of the HiResCAM on convolutional neural networks, Class Activation Mapping of shallow features was employed to automatically delineate the vertebrae regions. Subsequently, local contrast enhancement was conducted to improve the readability of the vertebral areas while avoiding the amplification of background noise. Pixel-level alpha blending was applied at the edges of the ROI to eliminate the boundaries between regions. Experimental results on real medical images demonstrate that this method significantly enhances the readability of the vertebral areas in lateral cervical spine images without damaging peripheral information, surpassing previous approaches with its human-like focus on the local ROIs.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
