PaperlayerView exact revision on arXiv

Source-linked AI summary

Domain Adaptation for Medical Image Analysis: A Survey

Hao Guan, Mingxia Liu

arXiv:2102.09508v1cs.CVcs.LGeess.IV

TL;DR

Medical image analysis is constrained by domain shift and scarce labeled data across datasets. This survey synthesizes medical-image domain adaptation methods, organizes them by model type and supervision, and summarizes benchmark datasets, challenges, and future directions.

  • Problem

    Medical image datasets often have limited labeled samples, while existing models may remain poorly adapted to medical images and differing imaging domains.

  • Method

    The paper surveys domain adaptation advances, categorizes methods into shallow and deep models with supervised, semi-supervised, and unsupervised subdivisions, and summarizes benchmark datasets.

  • Results

    The survey presents recent domain adaptation models across medical image analysis tasks and organizes their reported advances, challenges, and research directions.

  • Takeaways & Limitations

    The survey provides a structured overview intended to improve understanding of current domain adaptation studies and future directions in medical image analysis.

Abstract

from arXiv · show

Machine learning techniques used in computer-aided medical image analysis usually suffer from the domain shift problem caused by different distributions between source/reference data and target data. As a promising solution, domain adaptation has attracted considerable attention in recent years. The aim of this paper is to survey the recent advances of domain adaptation methods in medical image analysis. We first present the motivation of introducing domain adaptation techniques to tackle domain heterogeneity issues for medical image analysis. Then we provide a review of recent domain adaptation models in various medical image analysis tasks. We categorize the existing methods into shallow and deep models, and each of them is further divided into supervised, semi-supervised and unsupervised methods. We also provide a brief summary of the benchmark medical image datasets that support current domain adaptation research. This survey will enable researchers to gain a better understanding of the current status, challenges.

I. INTRODUCTION

Medical image analysis faces domain shift because source and target datasets often differ in distribution, scanners, protocols, and subject cohorts. This survey reviews domain adaptation methods and organizes them by model type, supervision, and related settings.

  • Test error generally increases as the distribution difference between training and test datasets grows, including for deep CNNs.
  • Medical imaging has limited labeled data because annotation is expensive, time-consuming, and requires expert participation.
  • Domain adaptation addresses heterogeneity among related medical image datasets by minimizing distribution differences between domains.
  • Prior surveys mainly covered natural-image domain adaptation or broad medical transfer learning, leaving limited focused reviews of medical-image domain adaptation.
  • The survey reviews recent methods, categorizing them into shallow and deep models, each subdivided into supervised, semi-supervised, and unsupervised approaches.
  • Domain adaptation transfers knowledge from a labeled source domain to a target domain for a shared task under different domain distributions.

C. Different Settings of Domain Adaptation

Domain adaptation methods can be classified by model type, label availability, modality difference, source count, and adaptation steps. These settings overlap, so the survey uses model type as its primary organizing criterion.

  • Model Type: Shallow DA uses engineered features and conventional models, whereas deep DA jointly learns features and task models end-to-end.
  • Label Availability: Supervised, semi-supervised, and unsupervised DA differ by the amount of labeled and unlabeled target data available during training.
  • Modality Difference: Single-modality DA uses the same modality across domains, while cross-modality DA adapts between different modalities.
  • Number of Sources: Single-source DA uses one source domain, whereas multi-source DA combines multiple source domains and must handle heterogeneity among them.
  • Adaptation Step: One-step DA directly adapts between source and target domains, while multi-step DA introduces intermediate domains when their relationship is relatively distant.
  • These categories are not mutually exclusive, and the survey organizes its medical-image review primarily around shallow versus deep models.

A. Overview

Shallow domain adaptation methods address domain shift by reweighting source instances or transforming source and target features into a shared space. The survey reviews these strategies and representative medical-imaging applications.

  • A. Overview: Shallow domain adaptation uses instance weighting and feature transformation as two common strategies.Instance weighting emphasizes source samples relevant to target samples, while feature transformation constructs a shared representation space.
  • A. Overview: Instance weighting assigns larger weights to source instances more relevant to target data before training a classifier or regressor.This reweighted training process is intended to reduce the source–target distribution gap.
  • A. Overview: Feature transformation maps source and target samples into a common representation space to reduce their distribution gap.Examples use nonlinear mappings, low-rank representation, or PCA before classifier training and target fine-tuning.
  • A. Overview: A two-step ultrasound adaptation framework first minimizes global domain differences with PCA, then fine-tunes a random forest using a few labeled target samples.The procedure combines shared-space transformation with limited target supervision.
  • A. Overview: Instance weighting improved classification on four chest CT datasets, while a feature-transformation method achieved good MCI conversion-prediction performance.The cited studies illustrate shallow adaptation across classification tasks and medical imaging datasets.

IV. DEEP DOMAIN ADAPTATION METHODS

Deep domain adaptation methods use CNN-based representations, fine-tuning, intermediate domains, task-specific architectures, and adversarial or semi-supervised objectives. The reviewed approaches span supervised, semi-supervised, and unsupervised settings.

  • IV. DEEP DOMAIN ADAPTATION METHODS: Deep domain adaptation builds on CNN features and includes supervised, semi-supervised, and unsupervised approaches.The survey discusses fine-tuning, task-specific 3D CNNs, reconstruction, and adversarial learning.
  • A. Supervised Deep DA: Using a small number of target training examples for fine-tuning can improve the transferability of pre-trained brain-lesion segmentation models.The finding concerns CNN models pre-trained on brain MRI scans.
  • A. Supervised Deep DA: Two-step adaptation achieved better results than direct transfer in skin-cancer classification on MoleMap and HAM10000.The method first fine-tunes on a relatively large skin-cancer dataset and then trains on the smaller target dataset.
  • A. Supervised Deep DA: Task-specific 3D CNNs are designed to exploit information in volumetric medical images for classification and segmentation adaptation.Examples pre-train on source-domain MR images and fine-tune upper layers or image patches on target data.
  • B. Semi-Supervised Deep DA: Semi-supervised methods combine labeled source data with unlabeled or labeled target data through reconstruction, fine-tuning, or three-category GAN discrimination.The Y-Net approach reconstructs source and target images before target-label fine-tuning.
  • C. Unsupervised Deep DA: Unsupervised deep adaptation avoids labeled target data and commonly learns domain-invariant features with DANN-like adversarial architectures.Domain discriminators are trained alongside task networks to support adaptation across imaging domains.

1) Feature Alignment:

Feature-alignment methods reduce domain differences within learned representations, often through adversarial losses or distribution-matching objectives. The survey also describes image-level translation and combined image–feature alignment for cross-modality and cross-vendor adaptation.

  • 1) Feature Alignment:: DANN-like models jointly train task networks and domain discriminators to learn domain-invariant features.This strategy is applied to segmentation and classification across multiple medical imaging domains.
  • 1) Feature Alignment:: Cross-modality cardiac segmentation can adapt only low-level layers when experiments support the assumption that modality shifts mainly affect low-level characteristics.The approach uses fixed higher layers and was evaluated on MM-WHS.
  • 1) Feature Alignment:: CMD-based adaptation regularizes higher-order central moments of feature distributions, with the loss imposed on fully connected layers for fMRI brain-activity classification.The method targets high-level features assumed to exhibit substantial domain shift.
  • 2) Image Alignment:: CycleGAN translates unpaired source-domain MR images into target-domain CT images and reconstructs the source image using cycle consistency.The reconstruction difference between the input and returned source image is minimized.
  • 2) Image Alignment:: Image-level translation supports adaptation for breast-cancer classification, OCT denoising, synthetic tumor generation, and noise-style transfer.These applications treat differences in medical centers, noise levels, or image styles as domain shifts.
  • 3) Image+Feature Alignment:: Combined image and feature alignment uses translated source images plus adversarial feature matching for cardiac segmentation across modalities or vendors.The reviewed vendor setting includes Philips, Siemens, and GE cine MRI data.

3) Image+Feature Alignment:

Further methods separate domain-invariant content from domain-specific style, or use consistency, correspondence, autoencoding, multi-domain, and batch-normalization mechanisms. These approaches target adaptation when domains share structure but differ in appearance or acquisition setting.

  • 4) Disentangled Representation:: Disentangled representation learning embeds each domain into shared content and domain-specific style spaces, then adapts in the invariant content space.The approach was evaluated for CT–MRI liver segmentation on LiTS and a local dataset.
  • 4) Disentangled Representation:: Self-ensemble adaptation compares student and teacher predictions with a consistency loss on unlabeled target images.The teacher is updated by exponential moving average of student weights, and related work adds an adversarial loss.
  • 4) Disentangled Representation:: Normalized cross-correlation identifies visually corresponding regions across electron-microscopy domains and converts high-scoring regions into soft target labels.The method exploits recurring regions of interest despite domain differences.
  • 4) Disentangled Representation:: A convolutional auto-encoder can transform AlexNet feature maps into medical-image-relevant features for subsequent classification.The adapter is trained as an alternative to relying only on labeled-target fine-tuning.
  • 4) Disentangled Representation:: Multi-domain adaptation uses adversarial losses with labeled source data and paired unlabeled target data, while lifelong adaptation fine-tunes domain-specific batch-normalization parameters.The latter shares convolutional filters across source domains but maintains separate BN parameters.
  • Benchmark datasets: The survey introduces benchmark medical imaging datasets organized by research object or organ to support comparison across domain-adaptation studies.ADNI is highlighted as an influential Alzheimer’s disease resource involving MRI, PET, and fMRI datasets.

1) ADNI:

The benchmark datasets span Alzheimer’s disease, autism, lesion segmentation, and brain tumor analysis, with substantial variation in imaging centers, scanners, modalities, and sample structures.

  • AIBL supports Alzheimer’s disease research with MRI and PET data collected from two medical centers in Australia.
  • CADDementia provides multi-center T1-weighted MRIs from patients with Alzheimer’s disease, mild cognitive impairment, and healthy controls.
  • IXI contains 600 MR images from healthy subjects collected at three London hospitals using different scanners.
  • ABIDE aggregates functional and structural brain imaging data from more than 24 imaging centers worldwide for autism research.
  • Brain-imaging benchmarks also include ISBI2015 MS lesion data, BraTS multi-center tumor MRIs, and WMH images acquired across five scanners and three institutes.

9) CE-MRI dataset:

The listed benchmarks cover multimodal brain research, chest imaging, cardiac cross-modality adaptation, and prostate or lung cancer studies, often reflecting multi-site acquisition.

  • HCP includes 1,200 subjects scanned in four modalities on a 3T scanner, with 200 additional subjects scanned on a 7T scanner.
  • Chest-imaging benchmarks include NIH ChestXray14 with 112,120 frontal-view X-rays across 14 disease categories, plus datasets from Indiana hospitals and clinical centers.
  • MM-WHS provides 20 MR and 20 CT cardiac images with pixel-wise annotations; unpaired images from different sites support cross-modality adaptation.
  • PLCO contains about 196,000 X-ray images acquired from 10 imaging centers throughout the United States.

3) NIH Chest:

The benchmarks cover retinal imaging, mammography, and histopathology, with datasets varying in size, labeling, acquisition format, and multi-center composition.

  • DRIVE contains 40 color retinal images with pixel-level ground truth, split evenly into training and test sets.
  • STARE provides 400 eye images with partially labeled blood-vessel segmentation data, while SINA contains 220 OCT B-scans from 20 eye volumes.
  • CBIS-DDSM includes 3,103 scanned mammograms labeled benign or malignant with verified pathology information.
  • InBreast contains mammograms from 115 patients with benign or malignant labels and BI-RADS readings, but lacks biopsy confirmation.
  • The CAMELYON challenge dataset contains histological images from 100 patients collected by five medical centers in the Netherlands.

3) CAMELYON:

The benchmarks include mammography, dermoscopy, and prostate MRI datasets, spanning large multi-source collections and smaller segmentation datasets acquired across hospitals.

  • MIAS contains 322 mammogram images with mediolateral oblique views, stored as scanned film copies at 1024 × 1024 resolution.
  • MoleMap contains 102,451 skin images covering 25 conditions, including three cancerous and 22 benign categories.
  • HAM10000 provides 10,015 multi-source dermatoscopic images across seven skin categories, including two cancerous and five benign categories.
  • ISIC houses 20,000 dermoscopic images in three skin-condition categories acquired from several centers.
  • PROMISE12 contains 50 T2-weighted prostate MRIs with segmentation ground truth acquired in different hospitals, while the BWH dataset contains MRIs from 15 subjects.

2) BWH:

The section introduces benchmark datasets used in medical image analysis, spanning liver segmentation, breast-cancer histology, colorectal tissue classification, disc localization, and spinal-cord MRI.

  • LiTS contains 201 CT scans for liver segmentation, with annotations independently reviewed by three radiologists.
  • NKI and VGH provide independently collected breast-cancer histology datasets containing 778 and 666 images, respectively.NKI includes 248 patients and VGH includes 328 patients; both use 1128 × 720 images.
  • The IHC dataset contains 1,377 colorectal-cancer tissue samples from 643 patients labeled as epithelium or stroma.
  • A public lower-spine dataset supports disc localization and segmentation using 16 multimodality MRIs with manual segmentations.
  • The SCGM dataset contains MRI data from 80 healthy subjects recruited across four medical centers.
  • Medical imaging datasets pose adaptation challenges because images are high-dimensional 3D or 4D data with extensive structural context across slices and time.Time-series fMRI may contain series of 3D volumes, each composed of hundreds of 2D slices.

1) 3D/4D Volumetric Representation:

The section identifies limited labels, modality and domain heterogeneity, and high-dimensional volumetric structure as major challenges for medical-image domain adaptation. It highlights unsupervised, cross-modality, and multi-source or multi-target adaptation as important research directions.

  • 1) 3D/4D Volumetric Representation:: Medical datasets often have limited samples and even fewer labels because expert annotation is expensive and time-consuming.Off-the-shelf ImageNet-pretrained models may not adapt well to medical images.
  • 1) 3D/4D Volumetric Representation:: Inter-modality heterogeneity complicates adaptation across CT, structural MRI, functional MRI, and PET despite their complementary information.Structural MRI and PET scans from the same subject can exhibit large inter-modality discrepancies.
  • 1) 3D/4D Volumetric Representation:: 3D models generally explore medical-image features more powerfully than ImageNet-pretrained 2D CNNs, while task-specific ROI definition remains an open issue.Few domain-adaptation studies develop 3D/4D models with task-specific ROI definitions.
  • 1) 3D/4D Volumetric Representation:: Unsupervised domain adaptation reduces reliance on labeled target data, while domain generalization and zero-shot learning avoid target data altogether.
  • 1) 3D/4D Volumetric Representation:: Cross-modality adaptation seeks models trained on one modality that generalize to another, while multi-modality settings require handling inter-modality and inter-domain differences.CycleGAN and disentangle learning have been introduced, but further exploration remains necessary.
  • 1) 3D/4D Volumetric Representation:: Multi-source and multi-target adaptation remains underexplored despite its clinical significance for multiple imaging centers and target domains.
  • 1) 3D/4D Volumetric Representation:: The survey organizes domain-adaptation methods into shallow and deep models, each divided into supervised, semi-supervised, and unsupervised categories, and summarizes benchmark datasets and future challenges.
Loading 2102.09508v1…