MiB: A class-incremental learning method for semantic segmentation in natural images.

CLIP-CT: A Continual learning method for abdominal multi-organ and tumour segmentation.

Saving100x: A class-incremental learning method for semantic segmentation in natural images that leverages prototype replay.

Adaptive Prototype: A class-incremental learning method for semantic segmentation in natural images that adaptively updates prototypes.

These methods cannot handle few-shot learning and domain shifts.

Domain-Incremental Learning:

MDIL: A class- & domain-incremental learning method for semantic segmentation in natural images that employs multiple decoders for different domains with a shared encoder.

MDIL uses a separate decoder per domain, which limits scalability as the number of domains increases. It cannot handle few-shot incremental classes.

Few-shot Class-Incremental Learning:

PIFS: A prototype-based method for few-shot class-incremental semantic segmentation.

Subspace: A regularization-based method for few-shot class-incremental learning.

C-FSCIL: A meta-learning method for few-shot class-incremental learning.

FACT: A few-shot class-incremental learning method that reserves the embedding space for new classes in the base session for future possible extensions.

NC-FSCIL: A few-shot class-incremental learning method that proposes to fix a learnable classifier as a geometric structure instructed by neural collapse. Zhong et al. showed that a semantic segmentation model with a classifier fixed as a simplex equiangular tight frame (ETF) performs significantly worse than a model with a learnable classifier.

Gen-Replay: A data-free replay method that implements Liu et al. with a diffusion model to generate 3D medical volumes.

GAPS: A few-shot class incremental learning method that utilizes guided copy-paste augmentation to synthesize diverse training data in semantic segmentation.

SoftNet: A few-shot class incremental learning method that jointly learns the model weights and adaptive soft masks to minimize catastrophic forgetting and to avoid overfitting novel few samples.

FSCIL-SS: A few-shot class incremental learning method that uses pseudo-labeling to augment novel classes and leverages knowledge distillation to prevent forgetting.

FeCAM: A few-shot class incremental learning method that investigates methods to enhance the representation of class prototypes in CIL, aiming to improve plasticity within the stability-favoring classifier-incremental setting. The method evaluates domain incremental learning as well, separately.

Although effective for class-incremental and few-shot learning, these methods are not robust to domain shifts.

Semi-Supervised based methods:

These methods can handle unlabeled data.

RETRIEVE: A coreset selection framework for efficient and robust semi-supervised learning.

NNCSL: A nearest-neighbor-based continual semi-supervised method.

UaD-CE: It uses a Class Equilibrium module to address overfitting and an Uncertainty aware Distillation module to distill reliable knowledge for memorizing previous categories and eliminate the ambiguity between previous and novel categories.

CSL: CSL proposes pseudo-label selection as a convex optimization problem within the confidence distribution. By doing so, CSL aims to overcome the challenges of overconfidence and context loss, ultimately enhancing the performance of semi-supervised semantic segmentation models.

These methods cannot jointly address domain shifts, class-incremental, and few-shot learning.

Others (Representation Learning, Few-shot Learning, Meta-Learning, Active Learning and Domain-shift):

SupCL, UnSupCL, and UnSupCL-HNM are contrastive learning methods designed for supervised, unsupervised, and hard negative mining settings, respectively.

MTL: MTL enables fast adaptation to unseen tasks through efficient training. It bridges gradient-based meta-learning and multi-task learning, making it particularly suitable for few-shot learning scenarios.

MAML: A multi-task representation learning framework that improves meta-learning via spectral-based regularization, ideal for few-shot learning.

CLIP-driven: A universal and extensible Language-Vision model for organ segmentation and tumor detection from abdominal computed tomography. It is a large pre-trained CLIP-driven U-Net model, which has already been exposed to most of the 35 classes in the Med FoSSIL-Mixed setting.

HALO: A hyperbolic neural network approach to pixel-level active learning for semantic segmentation under domain shift.