Brain Tumor Segmentation Using Multi-Task Learning

For cancer diagnosis and treatment planning, accurate and reliable brain tumor segmentation is crucial. In this paper, a CNN encoder-decoder-based multi-task learning method is proposed. This was designed by using a depth-shared encoder to extract useful features in-depth during down-sampling operations and having a segmentation decoder with long skip connections through additive layers that were introduced in ResNet, from the corresponding encoder spatial levels to do the segmentation tasks. A similar structure to that of a segmentation decoder, which includes an up-sampling layer with long skip connections through an additive layer from the corresponding encoder spatial levels, acts as a feature reconstruction decoder to regularize the shared encoder during training. The experimental results showed that the proposed methods were obtained with a dice similarity coefficient of 71.42%, intersection over union of 51.48%, sensitivity of 76.89%, and specificity of 99.87% in training. In addition, the injection of the Gaussian noise into the input results in increased generalizability of the model. We also compared the outcomes of multitasking learning with the single-task learning approach.