Table 1 The table below summarizes recent studies on MS lesion segmentation using CNNs.

Brown RA et al¹⁷

Own dataset

FCNN

Agreement with manual segmentation

Dice score: 0.74 (Jacard index)

Coronado I et al¹⁸

Own dataset

3D CNN

High false-positive rate in small lesions

Dice score: 0.77

Essa E et al¹⁹

MICCAI 2008 MS challenge dataset

Region-based Convolutional Neural Network (R-CNN)

Need for large annotated datasets

Dice score: 0.83

Birenbaum A et al²⁰

2015 Longitudinal MS Lesion Segmentation Challenge

Single View CNN (V-Net) and Longitudinal Network (L-Net)

Performance compared to trained human raters

Dice score: 0.627

Aslani S et al²¹

ISBI 2015, Private dataset

Deep end-to-end 2D CNN

Requires validation on larger datasets

Dice score: 0.6114 (ISBI), 0.6655 (Private)

Nichyporuk et al²². (2022)

Clinical trials datasets

Trial-conditioned CIN, naive pooling, single-trial baselines

Handling biases in the label generation process

Dice scores: 0.795,

Wiltgen et al²³

In-house dataset, MSSEG, ISBI 2015, MICCAI 2008

Ensemble of three 3D UNets

Requires large dataset for training, limited generalizability to unseen data

Dice score: 0.67

Gabr et al²⁴

CombiRx clinical trial dataset

FCNN

Variations in class sizes, reliance on multimodal MRI data

Dice scores: 0.95 (WM), 0.96 (GM), 0.99 (CSF), 0.82 (T2 lesions)

Duong et al²⁵

Hospital of the University of Pennsylvania

3D U-Net CNN

Variability in lesion characteristics and acquisition parameters

Dice score: 0.789,

Afzal et al²⁶

ISBI, MICCAI datasets

Cascaded 2D CNNs

Overlapping lesions, lesions near cortex

Dice scores: ISBI: 0.67, MICCAI: 0.72

de Oliveira et al²⁷

ISBI 2015, In-house dataset

FCNN

Limited test group size, need for larger validation

–