Abstract:Despite advances with deep learning (DL), automated airway segmentation from chest CT scans continues to face challenges in segmentation quality and generalization across cohorts. To address these, we propose integrating Curriculum Learning (CL) into airway segmentation networks, distributing the training set into batches according to ad-hoc complexity scores derived from CT scans and corresponding ground-truth tree features. We specifically investigate few-shot domain adaptation, targeting scenarios where manual annotation of a full fine-tuning dataset is prohibitively expensive. Results are reported on two large open-cohorts (ATM22 and AIIB23) with high performance using CL for full training (Source domain) and few-shot fine-tuning (Target domain), but with also some insights on potential detrimental effects if using a classic Bootstrapping scoring function or if not using proper scan sequencing.

What problem does this paper attempt to address?

The problems that this paper attempts to solve are: 1. **Improve the performance and generalization ability of airway segmentation**: Although deep learning (DL) has made progress in automatically segmenting airways from chest CT scans, it still faces challenges in terms of segmentation quality and cross - cohort generalization. Specifically, well - trained deep - learning models are often difficult to generalize to new cohorts, especially data obtained on different scanners or data from patients with lung diseases. 2. **Address the issue of expensive manual annotation**: In some cases, the cost of manually annotating a complete fine - tuning data set is too high. Therefore, this study focuses particularly on few - shot domain adaptation, that is, fine - tuning with only a small amount of annotated data on the target domain. To solve these problems, the authors propose to introduce curriculum learning (CL) into the training process of the airway - segmentation network. By dividing the training set into batches according to the complexity scoring function and gradually increasing the task difficulty, the model can first learn from simple examples and then handle more complex ones. This method aims to improve the performance of the model on the source domain and achieve better generalization on the target domain through few - shot domain adaptation. ### Specific contributions 1. **Introduce a new complexity scoring metric**: The authors propose a complexity scoring metric based on knowledge transfer, which combines the visual attributes of CT scans and the ground - truth segmentation to evaluate the complexity of lung CT scans. And it is compared with the bootstrapping scoring function. 2. **CL - supervised training framework**: A CL - supervised training framework is proposed to enhance the performance of the deep - learning baseline segmentation network on the source domain. 3. **CL - based few - shot incremental domain - adaptation method**: A CL - based few - shot incremental domain - adaptation method is proposed to fine - tune the best network trained on the source domain on the target domain. This helps the model retain the knowledge of the source domain while adapting to the target domain, reducing the risk of catastrophic forgetting. ### Method overview - **Complexity scoring function**: - **Bootstrapping scoring function**: Calculate the complexity score \( CS_B = 1-\text{IoU}_{\text{SEG1}} \) based on the segmentation results of the pre - trained network (SEG1). - **Machine - learning - based scoring function**: Use a random forest classifier to extract features from CT scans and their ground - truth segmentations to predict the segmentation quality, thereby generating a complexity score. - **CL training settings**: - **Full training**: Conduct CL training on the source data set and explore different batch combinations. - **Incremental domain adaptation**: Conduct few - shot fine - tuning on the target data set, gradually introduce target - domain samples to reduce catastrophic forgetting and improve generalization ability. Through these methods, the authors hope to improve the effect of airway segmentation on multiple performance indicators and solve the problems of cross - cohort generalization and high cost of manual annotation.

Curriculum Learning for Few-Shot Domain Adaptation in CT-based Airway Tree Segmentation

Domain Adaptation Meets Zero-Shot Learning: an Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation

Few-Shot Airway-Tree Modeling using Data-Driven Sparse Priors

Accurate Airway Tree Segmentation in CT Scans via Anatomy-aware Multi-class Segmentation and Topology-guided Iterative Learning

Multiscale Unsupervised Domain Adaptation for Automatic Pancreas Segmentation in CT Volumes Using Adversarial Learning.

Deep anatomy learning for lung airway and artery-vein modeling with contrast-enhanced CT synthesis

CADA: Multi-scale Collaborative Adversarial Domain Adaptation for unsupervised optic disc and cup segmentation

A scale-aware UNet++ model combined with attentional context supervision and adaptive Tversky loss for accurate airway segmentation

Contrastive Semi-Supervised Learning for Domain Adaptive Segmentation Across Similar Anatomical Structures

CXR Segmentation by AdaIN-based Domain Adaptation and Knowledge Distillation

CFDA: Collaborative Feature Disentanglement and Augmentation for Pulmonary Airway Tree Modeling of COVID-19 CTs

Continual Test-time Domain Adaptation via Dynamic Sample Selection

An end-to-end multi-scale airway segmentation framework based on pulmonary CT image

Joint semi-supervised and contrastive learning enables zero-shot domain-adaptation and multi-domain segmentation

Semi-Supervised Medical Image Segmentation Based on Deep Consistent Collaborative Learning

Curriculum-style Local-to-global Adaptation for Cross-domain Remote Sensing Image Segmentation

An Unsupervised Domain Adaption Framework for Aerial Image Semantic Segmentation Based on Curriculum Learning

Enhancing Bronchoscopy Depth Estimation through Synthetic-to-Real Domain Adaptation

Curriculum-Based Augmented Fourier Domain Adaptation for Robust Medical Image Segmentation

Collaborative Teaching with Attention Distillation for Multiple Cross-Domain Few-Shot Learning

A Multilevel-Guided Curriculum Domain Adaptation Approach to Semantic Segmentation for High-Resolution Remote Sensing Images.