Single-image high-dynamic-range reconstruction for intraoral dental x-ray images
Sunghoon Choi, Jin-Woo Jeong, Geun-Young Oh, Dong‐Il Kim, Yang Jing-yu, Yoon‐Ho Song · Medical Imaging 2022: Physics of Medical Imaging · 2022
Intraoral dental x-ray exams are the cost-effective way for early detection of decay between teeth. The scintillation-based x-ray detectors coupled with fiber optic plate (FOP) with CMOS sensor are well-suited for high-resolution intraoral radiography. However, the transmitted x-ray photons, which are not absorbed in the FOP, directly hit the end of CMOS sensors, thus results in the salt and pepper noise on the image. Another issue is the optimization of contrast and brightness when the foreign bodies exist in tooth. A standard tone-mapping operator (TMO) utilizes a full scale of 14-bit gray range so that a whole pixel values including the outlier (high density materials) are distributed over the fixed range. This would degrade the contrast of the soft tissue because the TMO is still limited to low dynamic range (LDR) and associated with the varying parameters for the pixel value correction. A deep neural network-based approach was proposed to expand the LDR image to synthetic high dynamic range (HDR) image in order to enhance the contrast of the soft tissue after the TMO. We have also conducted the adaptive median filtering to reject the salt and pepper noise on the images before the network training. We firstly added the salt and pepper noise on the open camera dataset (24-bit color RGB) and then validated our method using the periapical intraoral dental x-ray images. The results indicated that the noise-corrupted LDR images were optimally reconstructed into HDR images for both simulation and experiment dataset. The tone-mapped LDR from the synthetic HDR images showed the enhanced PSNRs by factors of 1.08 – 1.21 by comparing the unprocessed LDR images. In conclusion, the proposed parameter-free deep learning method may provide better outcomes in practical intraoral radiography.