Report on Current Developments in Tumor Lesion Segmentation in PET/CT Imaging
General Direction of the Field
The field of tumor lesion segmentation in Positron Emission Tomography (PET) and Computed Tomography (CT) imaging is rapidly advancing, driven by the need for more accurate and reliable diagnostic tools in oncology. Recent developments are particularly focused on improving the generalization and robustness of deep learning models across different tracers (such as Fluorodeoxyglucose (FDG) and Prostate-Specific Membrane Antigen (PSMA)) and clinical sites. This is crucial because variations in tracer uptake patterns, acquisition protocols, scanner types, and patient populations can significantly affect image quality and lesion detectability.
One of the key trends in the field is the use of ensemble models and multi-stage segmentation approaches. These methods aim to combine the strengths of various deep learning architectures to enhance segmentation accuracy and robustness. For instance, the integration of models like U-Net, SwinUNETR, and SegResNet in a staged manner has shown promise in refining lesion segmentation, particularly in complex datasets.
Another significant development is the exploration of novel preprocessing and normalization techniques. These techniques are designed to address the variability in PET/CT images, thereby improving the performance of segmentation models. For example, the introduction of periodic sine transformations (SineNormal) to PET data has been shown to enhance lesion detection by highlighting intensity variations and producing concentric ring patterns in PET highlighted regions.
Additionally, there is a growing interest in data-centric approaches, where the focus is on optimizing the dataset rather than just the model architecture. This includes strategies like "data dieting," where certain training samples are excluded based on their impact on model performance. This approach has demonstrated improvements in reducing false positives and enhancing overall segmentation accuracy.
Noteworthy Developments
ResEnc-Model Ensemble: A 3D Residual encoder U-Net model combined with preprocessing techniques and test-time augmentations achieved top performance in the Auto-PET III challenge, outperforming the baseline model with a Dice score of 0.9627.
Sine Wave Normalization: The SineNormal technique introduced periodic sine transformations to PET data, significantly improving segmentation accuracy, especially for multitracer PET datasets.
Data Diet Approach: Excluding easy samples from the training dataset led to a reduction in false negatives and improved Dice scores, demonstrating the potential of data-centric optimization in lesion segmentation.
