Comparative Study Between Single and Two-Step Approaches on Automatic Segmentation of Prostate Abnormalities Using Deep Learning

Accurate diagnosis through MRI is the main non-invasive method in the diagnosis of PCas through manual segmentation of the prostate lesion. This process is time-consuming and prone to interreader variability. Deep learning-based image processing has emerged as a solution to the problem. However, most deep-learning-based approaches in prostate lesion segmentation use a one-step direct lesion segmentation approach, which may cause the model to learn from unnecessary inputs that could contribute to noise. To address this problem, some researchers suggested introducing prior prostate anatomical knowledge using a two-step approach that involves segmenting the prostate gland and using the extracted prostate gland as the input for the lesion segmentation model. Despite demonstrating high performance in lesion segmentation, there remains a lack of systematic comparative studies that quantify its practical advantages using standardised metrics, consistent architectures, and the same dataset. In this study, a comparative evaluation of one-step and two-step approaches in prostate lesion segmentation was conducted using the benchmark dataset Prostate158 with DeepLabV3, U-net, TransUNet, and TransAttUnet. The findings reveal that all four models with the two-step approach achieve significantly higher results relative to the one-step approach, with a percentage increase of 160.0%, 105.4%, 100.6%, and 68.1% for DeepLabV3, U-net, TransUNet, and TransAttUnet, respectively. However, the sequential dependencies of the two-step approach may also introduce error propagation and inconsistent results.