Cone beam computed tomography enhancement using feature‐embedded variational autoencoder with a perceptual loss function

Image‐Guided Radiation Therapy (IGRT) is a cancer treatment method. IGRT usually follows a fragmented treatment track. The weight loss of a patient causes the first treatment plan to be flawed. Medical experts can avert retaking Fan Beam Computed Tomography (FBCT) images before each session of the therapy if the Cone Beam Computed Tomography (CBCT), which is mounted on a Medical Linear Accelerator (LINAC), gives high‐quality images. Rather than improvising the CBCT machine, deep‐learning techniques can be used to enhance the CBCT images. This paper proposes a Feature‐Embedded Variational AutoEncoder (FE‐VAE) for CBCT image enhancement. A feature embedding is incorporated using a MultiResUNet‐based feature extractor to preserve structural information in the input image. A perceptual loss function, which is the product of Mean Absolute Error (MAE) and inverse Structural Similarity Index Measure (SSIM), is proposed in this work. The model is trained using Head‐and‐Neck (H&N) CBCT‐FBCT pairs of 59 cancer patients. Our proposed architecture successfully generates images much closer to the ground truth FBCT images. The model gives an average peak signal‐to‐noise ratio of 32.89 dB, mean squared error of 297.36, MAE of 32.37 HU, and SSIM of 0.99. The trained model gave the least deviated SSIM and MAE, implying that the model is well‐optimised. The experimentation with different losses proves the prominence of the proposed loss function. The visual quality of the results obtained is also comparable with the ground truth. The results indicate that the model produces enhanced CBCT images. Hence, the proposed FE‐VAE is useful in fractionated radiotherapy to minimise the number of scanning sessions for replanning the dosimetry.