Investigating efficient CNN architecture for multiple sclerosis lesion segmentation

Abstract

Purpose: The automatic segmentation of multiple sclerosis lesions in magnetic resonance imaging has the potential to reduce radiologists' efforts on a daily time-consuming task and to bring more reproducibility. Almost all new segmentation techniques make use of convolutional neural networks, with their own different architecture. Architectural choices are rarely explained. We aimed at presenting the relevance of a U-net like architecture for our specific task and at building an efficient and simple model. Approach: An experimental study was performed by observing the impact of applying different mutations and deletions to a simple U-net like architecture. Results: The power of the U-net architecture is explained by the joint benefits of using an encoderdecoder architecture and by linking them with long skip connections. Augmenting the number of convolutional layers and decreasing the number of feature maps allowed us to build an exceptionally light and competitive architecture, the MPU-net, with only approximately 30,000 parameters. Conclusion:The empirical study of the U-net has led to a better understanding of its architecture. It has guided the building of the MPU-net, a model far less parameterized than others (at least by a factor of seven). This neural network achieves a human level segmentation of multiple sclerosis lesions on FLAIR images only. It shows that this segmentation task does not necessitate overly complicated models to be achieved. This gives the opportunity to build more explainable models which can help such methods to be adopted in a clinical environment.

The brainstem in multiple sclerosis: MR identification of tracts and nuclei damage

“…Deep learningbased approaches for brain MRI are gaining interest due to their self-learning and generalization ability over large amounts of data [45]. A variety of methods have been developed and applied in the context of MS, including identification of multiple sclerosis subtypes or automatic lesions segmentations [48][49][50].…”

CNN for multiple sclerosis lesion segmentation: How many patients for a fully supervised method?

“…We started with the MPU-net architecture see Fig. 1 but as said in [8], it can be improved especially by adding regularizers such as dropout [9] and batch normalization [10] layers. Inspired by the U-net++ [11] we wanted to add deep supervision and to extend the architecture as shown in Fig.…”

“…During the study, following [8], only T2-fluid-attenuated inversion recovery (FLAIR) images were used among all MR images available in each exam since they are the most effective in practice [14].…”

“…The ground truth segmentation of MS lesion from 4 different radiologists and 1 consensus are included. The split between training and validation sets have been performed accordingly to [8] with exams from one patient as the testing set and the exams from all remaining patients as training set.…”

“…One hypothesis advanced was that the data set size was already sufficient and, therefore, pre-training would not improve segmentation predictions. Driven by previous work on finding an efficient CNN architecture for MS lesion segmentation in terms of number of learnable parameters the Minimally Parameterized U-net (MPU-net) [8], we propose 1) to improve the MPU-net and 2) to assess the needed number of training patients for this light architecture. We dit not find published architectures with a comparable low number of learnable parameters.…”