Uncertainty quantification in ultrasound image reconstruction

Ultrasound imaging is one of the most commonly used modalities in clinical practice due to its non-invasive nature and the absence of ionizing radiation. Such images are typically produced with well-established algorithms for converting acoustic signals into visual outputs. In recent years, artificial intelligence has experienced remarkable growth in the field of medical imaging, particularly through the application of deep learning techniques. These models have shown high performance and the potential to transform the field with novel methods for image reconstruction from raw data. However, their native lack of explainability is a major limitation, and issues regarding their reliability and clinical uptake are raised. As such, there has been increasing effort in recent years to develop uncertainty quantification metrics that can complement medical decision-making by providing insight into the confidence of model predictions. The aim of this thesis was to investigate the possibility of defining an uncertainty measure that can quantify the confidence of image reconstructions performed by a deep learning model on ultrasound data to create a measure that would enable clinicians and scholars to establish the reliability of AI-produced reconstructions. The employed database was based on a collection of 8,009 raw ultrasound data consisting of 8,000 simulated samples and 9 real ultrasound acquisitions. The reconstruction was performed based on a modified U-Net network consisting of two independent decoder branches that were designed to perform image reconstruction and anatomical segmentation simultaneously from raw ultrasound signals. Monte Carlo Dropout was used to generate multiple stochastic reconstructions by repeating inference with dropout active. The dual-output network architecture of the network was utilized to develop an uncertainty measure that puts greater importance on the areas determined to be more relevant by the model by assigning greater value to the uncertainty in the segmented ares. This resulted in developing a new measure called Median of Weighted Uncertainty on the Reconstruction (MWUR). The uses of the measure as a loss for better reconstruction in network training were also examined. The network obtained a mean absolute error (MAE) of 0.0693 ±0.0130 when operating with its baseline loss function and 0.0684 ± 0.0123 when trained on the MWUR-based loss on simulated data. On real data, the corresponding MAE was 0.1189 ± 0.0289 and 0.1305 ± 0.0334. The MWUR measure proved to be effective in discriminating between inferences drawn from simulated data and actual data, and also in showing a correlation with mean absolute error. The measure has proven to be promising to quantify the uncertainty of ultrasound images reconstructed directly from raw signals by deep learning networks, for different configurations. However, the MWUR measure is based on segmentation results, limiting its usage to some network structures. Additionally, the two branch network requires more extensive training time. Despite these limitations, MWUR has the potential to enhance uncertainty estimation in deep learning based ultrasound image reconstruction and by providing a weighted measure focused on relevant regions, it offers more informative feedback on the reliability of reconstructed images. The results obtained with this metric have been promising, and future work could focus on training improvements to address its limitations and on developing a new, more flexible metric.