Role of Deep Learning in Automated Tumor Detection and Segmentation in MRI and CT Imaging

Abstract

Manual delineation of the tumour in magnetic resonance imaging (MRI) or computed tomography (CT) is time consuming, operator dependant and prone to considerable inter-observer variation. Although deep learning (DL) is a potential approach to automatic tumor detection and segmentation, there is still a lack of quantitative comparison between architectures and imaging modalities. In this work, five DL architectures – U-Net, Attention U-Net, ResUNet++, nnU-Net, and Transformer-based Swin-UNETR – were compared and evaluated for the two segmentation tasks: brain tumor segmentation (BraTS 2021 MRI subset) and pulmonary nodule/tumor detection and segmentation (LIDC-IDRI CT subset). The total number of MRI and CT volumes were split into training (70%), validation (15%) and testing (15%). All pre-processing steps included bias-field correction, resampling, intensity normalization and patch extraction. Five-fold cross validation was used to train the models on the Adam optimizer with dice–cross-entropy as the loss function. The Dice similarity coefficient (DSC), Intersection-over-Union (IoU), 95th-percentile Hausdorff distance (HD95), sensitivity, specificity and inference time were used to assess performance. The paired t test and one-way ANOVA models were used at the α = 0.05 level of significance. The classical region growing/thresholding methods were significantly outperformed by all DL models (p < 0.001). nnU-Net achieved the highest DSC for MRI (0.912 ± 0.021) and CT (0.887 ± 0.026), followed by Swin-UNETR. The transformer and attention models improved the accuracy of the boundaries of heterogeneous and small lesions. In general, nnU-Net achieved the highest accuracy-efficiency tradeoff, which enabled the incorporation of DL into the workflow of radiologists without replacing them in the diagnosis process.