MRI Multitask Learning (MTL)#

In MTL reconstruction & segmentation, the segmentation task can be performed either in an multiclass or a multilabel manner. The segmentation_mode needs to be set either to multiclass or multilabel to ensure the chosen configurations aligns with the segmentation approach. Make sure to add the background_class when multiclass is selected.

Multiclass vs. Multilabel Segmentation#

Feature

Multiclass

Multilabel

Pixel Constraint

Each class is dependent
(Only one class can be assigned per voxel)
Each class is independent
(Multiple classes can be assigned per voxel)

Channel Output (\(C\))

\(N + 1\)
(Includes explicit Background class)
\(N\)
(Background is implicit)

Activation Function

Softmax
\(\frac{e^{z_i}}{\sum e^{z_j}}\) (Coupled probabilities)
Sigmoid
\(\frac{1}{1 + e^{-z_i}}\) (Independent probabilities)

Background Logic

Explicit Class
\(P(BG) = 1 - \sum P(Foreground)\)
Implicit Absence
All channels \(\approx 0\)

Inference Decision

Argmax
(Select index with highest probability)
Thresholding
(Select all indices where \(P > 0.5\))

Target Label Format

Flat Integer Map \((H, W)\)
(or One-Hot)
Stacked Binary Masks \((H, W, N)\)

Loss Function

Categorical Cross-Entropy
(Penalizes the target class vs. all others)
Binary Cross-Entropy
(Penalizes each channel independently)

Probability Space

Joint Distribution (Sum = 1)

Independent Bernoulli Distributions