Benchmarking Deep Learning-Based Reconstruction in Photoacoustic Computed Tomography with Clinically Relevant Synthetic Datasets

Deep learning (DL)-based image reconstruction methods for photoacoustic computed tomography (PACT) have developed rapidly. However, comparisons among DL-based methods are hindered by using various datasets across publications. To address this challenge, open-source, realistic synthetic datasets with varying complexities are proposed for benchmarking DL-based acoustic inversion methods in PACT. The datasets contain over 8,000 objects generated from 1,500 pairs of stochastic three-dimensional healthy and lesion-inserted numerical breast phantoms, with variations in lesion locations. Image formation and data acquisition are simulated in three dimensions and two dimensions, respectively. Leveraging the datasets, a preliminary benchmarking study was conducted to assess the performance of both DL-based and model-based image reconstruction methods to compensate for acoustic aberrations in PACT. The assessment included qualitative comparisons and quantitative analysis using traditional image quality (IQ) metrics. Results demonstrate that despite achieving favorable IQ scores, DL-based reconstructions can fail to recover lesions, indicating limitations of these metrics for lesion estimation. Future work will include transducer response modeling and further investigations into typical reconstruction challenges. This benchmark study is expected to guide future efforts toward enhancing the effectiveness and clinical applicability of DL-based methods in PACT.