Improving automated multiple sclerosis lesion segmentation with a cascaded 3D convolutional neural network approach (1702.04869v1)

Abstract: In this paper, we present a novel automated method for White Matter (WM) lesion segmentation of Multiple Sclerosis (MS) patient images. Our approach is based on a cascade of two 3D patch-wise convolutional neural networks (CNN). The first network is trained to be more sensitive revealing possible candidate lesion voxels while the second network is trained to reduce the number of misclassified voxels coming from the first network. This cascaded CNN architecture tends to learn well from small sets of training data, which can be very interesting in practice, given the difficulty to obtain manual label annotations and the large amount of available unlabeled Magnetic Resonance Imaging (MRI) data. We evaluate the accuracy of the proposed method on the public MS lesion segmentation challenge MICCAI2008 dataset, comparing it with respect to other state-of-the-art MS lesion segmentation tools. Furthermore, the proposed method is also evaluated on two private MS clinical datasets, where the performance of our method is also compared with different recent public available state-of-the-art MS lesion segmentation methods. At the time of writing this paper, our method is the best ranked approach on the MICCAI2008 challenge, outperforming the rest of 60 participant methods when using all the available input modalities (T1-w, T2-w and FLAIR), while still in the top-rank (3rd position) when using only T1-w and FLAIR modalities. On clinical MS data, our approach exhibits a significant increase in the accuracy segmenting of WM lesions when compared with the rest of evaluated methods, highly correlating ($r \ge 0.97$) also with the expected lesion volume.

• The paper demonstrates a cascaded dual-CNN design that first targets sensitivity and then minimizes false positives.
• The method processes 3D patches to incorporate spatial context, achieving top performance on the MICCAI2008 dataset and private clinical data.
• Results include significant improvements in TPR, FPR, and DSC, suggesting robust clinical applicability for MS lesion segmentation.

Improving Automated Multiple Sclerosis Lesion Segmentation with a Cascaded 3D Convolutional Neural Network Approach

The paper presents an innovative approach to automate the segmentation of white matter (WM) lesions in multiple sclerosis (MS) patient images, leveraging a cascaded architecture of 3D convolutional neural networks (CNNs). This methodology addresses the inherent challenges of segmentation in brain MRI, including limited labeled datasets and the high imbalance between lesion and non-lesion tissue.

Methodology

The proposed method adopts a cascade of two 3D patch-wise CNNs. The first network targets candidate voxels with potential lesions, focusing on sensitivity, while the second network aims to reduce false positives by refining the initial predictions. This two-step approach enables a robust lesion detection by processing voxels of interest through successive, targeted inferencing, thereby enhancing overall segmentation performance.

Each CNN is trained independently, focusing on different sets of features. The design benefits from utilizing 3D patches that incorporate spatial contextual information pertinent to WM lesions, exploiting the volumetric nature of MRI data. Moreover, the architecture is maintained compact with a 7-layer configuration, optimizing for smaller datasets where overfitting might have been a concern with deeper networks.

Experimental Evaluation

The approach was evaluated using the MICCAI2008 MS lesion segmentation challenge dataset and two other private clinical datasets. On the MICCAI2008 dataset, the proposed method achieved the highest accuracy among 60 competing methods when all modalities (T1-w, T2-w, and FLAIR) were utilized. When evaluated on the private datasets, the method consistently outperformed established tools like LST and SLS, demonstrating superior segmentation accuracy and lesion volume correlation (r ≥ 0.97).

Results and Implications

The paper reports significant numerical improvements in crucial metrics such as TPR, FPR, and DSC, emphasizing the method's efficacy in both sensitivity and precision. A notable aspect is its adaptability, performing reliably across different datasets and acquisition protocols, which suggests considerable potential for broader clinical applications without extensive re-training.

The simplicity and effectiveness of the cascaded CNN design cater well to the typical constraints of medical imaging datasets. It efficiently utilizes limited labeled images, highlighting a pathway for future endeavors in medical imaging segmentation where data scarcity is a commonplace challenge.

Future Directions

This work opens avenues for integrating deep learning-based segmentation in routine clinical workflows, potentially enhancing early diagnosis and longitudinal disease monitoring. Future research can explore extending this architecture to accommodate more dynamic brain imaging modalities or adapting it to other lesion or tumor detection tasks in medical imaging.

In conclusion, this paper provides a tangible advance in using CNNs for MS lesion segmentation, presenting a framework that could significantly impact clinical practices and biomedical research in neuroimaging.