Improving Liver Tumor Segmentation Robustness with Physics -Informed Regularization of a ResNet50 Network

The clinical process of liver cancer diagnosis, treatment planning, and monitoring depends on the precise segmentation of liver tumors using 'Computed Tomography (CT) scans'. The segmentation of medical images has advanced significantly, thanks to deep learning techniques, particularly 'Convolutional Neural Networks', or CNNs. However, data-driven models by themselves might produce results that are not anatomically reasonable or that do not sufficiently account for known picture properties. This study investigates the application of Physics-Informed Neural Networks (PINNs) to enhance the segmentation performance of a ResNet50-based model in order to get beyond these limitations. By including physics-based constraints into the training process, we hope to direct the network to produce liver tumor segmentations that are more reliable and clinically significant. Our method uses a ResNet50 architecture to generate a baseline segmentation model, which is subsequently adjusted with a bespoke loss function. With a focus on encouraging uniformity in intensity inside the segmented regions and smoothness in the segmented tumor boundaries, this physics-informed loss is intended to penalize departures from expected image characteristics. Our PINN-enhanced model was tested on a specific test dataset against the baseline ResNet50. The quantitative findings show that the PINN technique effectively classifies pixels at the pixel level, achieving a high Accuracy of 97.15%, a robust F1-Score of 95.74%, and good Specificity of 100.00%. While the Dice Coefficient of 32.85% and Mean Intersection over Union (IoU) of 32.38% suggests that there could still be difficulties in obtaining precise region overlap, the low 'Mean Absolute Error (MAE)' of 0.0288 indicates that pixel predictions are generally correct.