Geometric Loss For Deep Multiple Sclerosis Lesion Segmentation

Abstract

Multiple sclerosis (MS) lesions occupy a small fraction of the brain volume, and are heterogeneous with regards to shape, size and locations, which poses a great challenge for training deep learning based segmentation models. We proposed a new geometric loss formula to address the data imbalance and exploit the geometric property of MS lesions. We showed that traditional region-based and boundary-aware loss functions can be associated with the formula. We further develop and instantiate two loss functions containing first-and second-order geometric information of lesion regions to enforce regularization on optimizing deep segmentation models. Experimental results on two MS lesion datasets with different scales, acquisition protocols and resolutions demonstrated the superiority of our proposed methods compared to other stateof-the-art methods.

QSMRim-Net: Imbalance-Aware Learning for Identification of Chronic Active Multiple Sclerosis Lesions on Quantitative Susceptibility Maps

“…2). Future work involves pairing QSMRim-Net with an automated T2-FLAIR lesion segmentation algorithm, such as All-Net [21] with geometric loss [54] and attention-based approaches [55] [56], followed by an automated method to separate confluents lesions [57]. We plan to adapt and train the algorithm to work directly on T2-FLAIR lesion segmentations.…”

“…of the lesions in this study were identified as rim+ lesions, posing a great challenge to learning based methods. We proposed DeepSMOTE for data oversampling to alleviate the data class imbalance, but as future work we plan to develop techniques on imbalance-aware loss functions, such as geometric loss [54]. A further limitation is inter-rater variability in identifying rim+ lesions.…”

Ensemble CNN and Uncertainty Modeling to Improve Automatic Identification/Segmentation of Multiple Sclerosis Lesions in Magnetic Resonance Imaging

“…Future directions of exploration will be: 1) the usage of specific pre-processing strategy for the FLAIR images, such that proposed in [51], to further improve the robustness of the method with respect to MRI and MRI scanners variability; 2) to check if the use of multiple imaging sequences could really contribute to a performance drop more than to the performance improvement; 2) the usage of a different loss function, such that proposed in [76], for better dealing with the problem of class unbalancing; 3) to study a "soft" consensus based on a single class (lesion) with different probability values, similarly to that proposed in [27], to reduce problem complexity; 4) related to the previous point #3, to explore a "soft" loss function, similar to [35], to better deal with "soft" consensus; 5) to use, besides FLAIR images, also complementary imaging modalities such as MPRAGE and MP2RAGE, to identify/segment, besides Draft WM lesion, also cortical lesions, similarly to [55]; 6) to test the proposed framework also for temporal MS progression (in this case, besides FLAIR, other imaging modalities, such as T 1 -w, should be used to define the lesions status); 7) referred to the previous point #6, to explore specific pre-processing strategies to make all the used modalities robust with respect to MRI and MRI scanners and to avoid drops in performances when using different data set [44].…”

“…CNN, compared to machine learning approaches, achieved remarkable success in biomedical image analysis [11,57,68]. DLM train and learn themselves to design features directly from data [6] and provide best results in several problems, including the case of MS lesion identification/segmentation [4,27,33,37,71,76]. This has also been confirmed in recent reviews [19,21,37,75].…”

DeDA: Deep Directed Accumulator

“…The state-of-the-art performance for object detection in 3D point clouds has also been achieved by network-predicted votes [54,55]. Memory U-Net [74] generates Hough votes using CNNs for lesion instance segmentation. This line of works mainly uses classical or learning models to generate votes for object detection or segmentation, leaving much room for exploiting the global prior information from the accumulator space.…”