Effective Loss Function for Unbalanced Breast Thermal Image Segmentation

The convolutional neural network's ability to learn images has reigned in computer vision tasks of object detection, classification, and segmentation. In segmentation, the CNN architectures of U-Net and SegNet have shown a good performance; thus, we implemented these networks to take the region of interest (ROI) of the breast thermal images from database for mastology research (DMR). We fine-tuned the networks to find the optimal hyperparameter that can result in the best learning performance. The networks were trained using the stochastic gradient descent optimization algorithm, and weights were updated using the error backpropagation of the algorithm. To minimize the error, the loss function is applied to evaluate a candidate solution of the weights. Thus, we conducted a preliminary study on finding the optimal loss function for breast thermal image segmentation. We applied loss functions of cross-entropy, dice, and focal and figured out the one that provides the highest segmentation accuracy. Then, to evaluate the segmentation result, we calculated the pixel accuracy rate and the dice coefficient to obtain the overlapping index measuring the overlay area between ground truth and predicted output. The result shows that the SegNet model with cross-entropy loss function can take the ROI of breast thermal images at the highest pixel accuracy and optimum dice coefficient of 0.8845 and 0.7928, respectively.