Handling Missing Data in Limited-View Photoacoustic Tomography Using Compressive Sensing Algorithm-Based Deep Learning

In photoacoustic tomography, missing sensor data poses a significant challenge for accurate image reconstruction. This study aims to evaluate various imputation methods to address different missing data scenarios, including random missing data, communication loss, and environmental interference. We applied mean imputation, median imputation, k-nearest neighbors (KNN), and multivariate imputation by chained equations (MICE) to these scenarios and assessed their performance using metrics such as MS-SSIM, SSIM, PSNR, R-squared, NMSE, and Entropy. Our results demonstrate that the presence of missing data significantly degrades image quality, with MICE consistently providing the best reconstruction performance across all scenarios. In random missing data cases, MICE achieved the highest SSIM and PSNR values, closely approximating the no missing data scenario. Similar trends were observed in communication loss and environmental interference cases, where MICE outperformed other imputation methods, followed by mean and median imputation, with KNN generally showing the lowest performance. These findings suggest that MICE is a robust method for handling missing data in PAT, enhancing image reconstruction quality.