Deep Attention-based Seminal Segmentation: A Practical Deep Learning Framework for Accurate Segmentation of the Hippocampus from Magnetic Resonance Images

Accurate segmentation of the hippocampus head and body from MR images is routinely performed to examine the link between the hippocampus deformation and neurological diseases, such as Alzheimer's and epilepsy. State-of-the-art seminal segmentation methods (including hippocampus segmentation) are based on deep learning algorithms. Most studies focused on developing robust deep learning algorithms to achieve satisfactory performance in hippocampus body and head segmentation. A critical aspect that has been overlooked in these studies is the strategy adopted to train deep learning algorithms when there are two or more target structures. In this work, we examine which deep learning training strategies would be more effective. These strategies include simultaneous (parallel segmentation of the head and body), serial (first head and then body), independent, and attention-based training and segmentation of the target structures. To this end, the hippocampus dataset from the Decathlon challenge and a residual neural network (Resnet) were employed to compare the above-mentioned strategies for hippocampus head and body segmentation. The Dice similarity coefficient and Hausdorff distance were calculated for the outcome of each strategy versus the manually defined hippocampus head and body masks. The quantitative analysis of the outcomes of different training frameworks demonstrated the superior performance of the attention-based training framework with Dice index of 0.89±0.03 (body) and 0.88±0.04 (head) compared to simultaneous, serial, and independent training frameworks with Dice indices of 0.88±0.04 (body) and 0.87±0.04 (head), 0.88±0.04 (body) and 0.87±0.04 (head), and 0.88±0.04 (body) and 0.86±0.04 (head), respectively. The statistical analysis demonstrated the significantly superior performance of the attention-based training framework (p-value<0.0001). In conclusion, the attention-based training framework is recommended for multi-structure seminal segmentation.