Investigation of Weighted Binary Cross Entropy and Resnet-50 on Segmentation Performance of Pleural Mesothelioma on CT Scans
Abstract
Purpose
This study addressed class imbalance in automated pleural mesothelioma segmentation and demonstrated that addressing this imbalance significantly improves agreement in the optimal probability map threshold, enabling simultaneous optimization of both spatial overlap and volume agreement metrics.
Methods
A dataset consisting of 5230 axial sections from computed tomography (CT) scans of 126 patients with pleural mesothelioma were retrospectively collected. For each scan, a single axial series was selected to ensure consistency in slice thickness and reconstruction filter. A radiologist contoured regions containing tumor, which served as the reference standard. Automated tumor segmentation was performed using a two-dimensional U-Net–based deep convolutional neural network (CNN), consisting of a symmetric encoder–decoder architecture with a pretrained VGG19 or ResNet50 backbone and corresponding up-sampling paths. The network produced pixelwise probability maps matching the input image resolution (512 × 512 pixels). Weighted binary cross-entropy (WBCE) was employed as a loss function. The results of this model were compared with those of the same model using a standard binary cross-entropy (BCE) loss function. A separate comparison was made with a model using a ResNet-50 backbone and WBCE. Segmentation performance was evaluated using (1) the Dice similarity coefficient (DSC) to quantify spatial overlap and (2) the fractional volume difference to assess volumetric agreement. Mean DSC and absolute volume difference were plotted as functions of network output threshold values for each model comparison.
Results
The WBCE-trained model simultaneously achieved a DSC of 0.68 and a fractional volume difference of 0.16, outperforming both previously published results and models trained using conventional BCE loss.
Conclusion
These results indicate that WBCE improves probability map calibration, supporting more reliable clinical deployment of CNN-based mesothelioma segmentation. Future studies will use WBCE in lieu of conventional BCE and investigate methods to improve model performance.