PaperlayerView exact revision on arXiv

Source-linked AI summary

Collaborative Unsupervised Domain Adaptation for Medical Image Diagnosis

Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, Mingkui Tan

arXiv:2007.07222v1cs.CVeess.IV

TL;DR

Medical image diagnosis is hindered by limited labels, domain discrepancies, and noisy source annotations. The paper proposes CoUDA, which uses peer networks for transferability-aware adaptation and collaborative label-noise handling, with theoretical and empirical evidence supporting its effectiveness and generalization.

  • Problem

    Medical UDA must learn from limited target labels while addressing source label noise and unequal sample transferability, assumptions commonly ignored by existing methods.

  • Method

    CoUDA uses diverse peer networks to weight hard-to-transfer samples, adapt noisy labels through a shared noise co-adaptation layer, and perform domain adaptation.

  • Results

    Extensive experiments on real-world medical and general images demonstrate CoUDA's effectiveness and generalization.

  • Takeaways & Limitations

    CoUDA provides a collaborative UDA approach that simultaneously addresses transferability differences and source label noise.

Abstract

from arXiv · show

Deep learning based medical image diagnosis has shown great potential in clinical medicine. However, it often suffers two major difficulties in real-world applications: 1) only limited labels are available for model training, due to expensive annotation costs over medical images; 2) labeled images may contain considerable label noise (e.g., mislabeling labels) due to diagnostic difficulties of diseases. To address these, we seek to exploit rich labeled data from relevant domains to help the learning in the target task via {Unsupervised Domain Adaptation} (UDA). Unlike most UDA methods that rely on clean labeled data or assume samples are equally transferable, we innovatively propose a Collaborative Unsupervised Domain Adaptation algorithm, which conducts transferability-aware adaptation and conquers label noise in a collaborative way. We theoretically analyze the generalization performance of the proposed method, and also empirically evaluate it on both medical and general images. Promising experimental results demonstrate the superiority and generalization of the proposed method.

I. INTRODUCTION

Medical image diagnosis lacks abundant clean annotations, motivating UDA that transfers knowledge from labeled source domains to unlabeled targets despite domain discrepancies. CoUDA addresses sample-level transferability differences and source label noise collaboratively with peer networks, supported by theoretical analysis and experiments across medical and general images.

  • Motivation: Medical image diagnosis often lacks rich supervision because expert annotation is costly, motivating UDA with labeled source and unlabeled target data.UDA learns domain-invariant features so a classifier trained on labeled source examples can apply to unlabeled target examples.
  • Challenges: Histopathological images exhibit domain discrepancies in scope shades, background colors, field views, appearance, and lesion-region scales.These differences can arise from tissue collection, sectioning, and staining procedures.
  • Challenges: Target samples differ in transferability, so hard-to-transfer examples require different alignment treatment from samples more similar to the source domain.Existing UDA methods commonly assume samples are equally transferable.
  • Challenges: Noisy source annotations caused by diagnostic difficulty and subjective bias can produce biased classifiers that perform poorly on the target domain.This remains problematic even when the target domain is aligned well with the source.
  • Proposed approach: CoUDA uses two peer networks to detect transferability, emphasize hard-to-transfer examples, denoise labels through a shared co-adaptation layer, and maximize classifier diversity.The method jointly conducts transferability-aware adaptation and label-noise handling.
  • Evidence: The paper analyzes CoUDA's generalization using Rademacher complexity and evaluates it on medical and general image classification tasks.The reported experiments demonstrate the proposed method's superiority and generalization.

A. Overall Collaborative Scheme of CoUDA

CoUDA uses multiple peer networks to jointly address transferability differences and label noise in unsupervised domain adaptation. Its training combines adaptive alignment, noise adaptation, and classifier diversity.

  • A. Overall Collaborative Scheme of CoUDA: CoUDA constructs two peer networks with shared domain discriminator and noise co-adaptation layer to address domain discrepancy and label noise collaboratively.The peer networks have separate feature extractors and classifiers but share the domain discriminator and noise co-adaptation layer.
  • A. Overall Collaborative Scheme of CoUDA: Collaborative domain adaptation detects sample transferability and applies a transferability-aware adversarial domain loss to align source and target features.
  • A. Overall Collaborative Scheme of CoUDA: Collaborative noise adaptation uses a noise co-adaptation layer and focal classification loss to improve robustness to noisy labels and class imbalance.
  • A. Overall Collaborative Scheme of CoUDA: Classifier diversity maximization keeps peer networks diverse, supporting transferability detection and collaborative noise handling.
  • A. Overall Collaborative Scheme of CoUDA: The overall training procedure is formulated as a minimax problem with trade-off parameters for domain and diversity losses.The parameters α and η control the domain loss and diversity loss, respectively.

B. Collaborative Domain Adaptation

CoUDA adapts domains by emphasizing samples whose transferability is difficult to assess and models noisy labels collaboratively. Focal classification loss further addresses class imbalance during noisy-label training.

  • B. Collaborative Domain Adaptation: Medical target images differ in domain-alignment difficulty, but existing UDA methods often treat all samples equally.These differences arise from variation in image appearance and lesion scale linked to inconsistent tissue preparation.
  • B. Collaborative Domain Adaptation: CoUDA estimates transferability from prediction inconsistency between two peer classifiers, using cosine distance to define a sample weight.Hard-to-transfer samples are associated with difficulty being correctly classified by both networks simultaneously.
  • B. Collaborative Domain Adaptation: The domain discriminator distinguishes source and target feature representations using transferability-aware domain loss, while feature extractors learn domain-invariant features adversarially.Source and target samples receive transferability-aware weights in the domain loss.
  • B. Collaborative Domain Adaptation: The noise co-adaptation layer estimates transitions from true labels to noisy labels using predictions and features, with two peers helping adjust single-network estimation errors.
  • B. Collaborative Domain Adaptation: Focal classification loss emphasizes minority classes, enabling CoUDA to handle class imbalance while training with noisy source labels.The focal-loss parameter is set to γ=2, and the loss uses predicted noisy labels for source data.

D. Classifier Diversity Maximization

CoUDA maximizes classifier diversity so peer networks remain useful for transferability detection and noise adaptation. The paper also describes ensemble inference and a Rademacher-complexity generalization analysis.

  • D. Classifier Diversity Maximization: Classifier diversity is maximized because transferability detection depends on classifier diversity and diversity prevents noise adaptation from collapsing to a single-network function.The diversity loss is defined using Jensen-Shannon divergence.
  • D. Classifier Diversity Maximization: The final CoUDA prediction averages the two networks’ predictions, and maximizing diversity can further improve ensemble performance.
  • D. Classifier Diversity Maximization: CoUDA implements adversarial optimization with a gradient reversal layer, allowing end-to-end training through standard backpropagation.
  • D. Classifier Diversity Maximization: Inference computes predictions from both peer networks, averages them, and returns the ensemble prediction.
  • D. Classifier Diversity Maximization: The generalization analysis uses Rademacher complexity and states that reducing transferability-aware domain discrepancy helps source classifiers match the target optimum.The theorem assumes an invertible and known noise transition matrix and a loss bounded in [0, M].
  • D. Classifier Diversity Maximization: The analysis assumes that a hypothesis performing well on both domains exists, because adaptation cannot proceed when no such hypothesis exists.

V. EXPERIMENTAL RESULTS

The experiments compare CoUDA with weakly supervised, unsupervised domain adaptation, and noisy-label baselines under a shared MobileNet V2 backbone. Evaluation includes four metrics on colon cancer diagnosis.

  • V. EXPERIMENTAL RESULTS: CoUDA is compared with MentorNet, Co-Teaching, NAL, DDC, DANN, MCD, CLAN, TCL, and a source-only variant using the same backbone.
  • V. EXPERIMENTAL RESULTS: Medical-image models are trained from scratch, while general-image experiments use ImageNet-pretrained MobileNet; training uses Adam with batch size 16 on one GPU.
  • V. EXPERIMENTAL RESULTS: The colon cancer diagnosis comparison reports Accuracy, Macro Precision, Macro Recall, and Macro F1-measure.

B. Evaluation on Colon Cancer Diagnosis

Colon cancer experiments evaluate CoUDA on noisy, imbalanced histopathology data, showing effective adaptation and contributions from transferability-aware and noise-adaptation components.

  • Dataset: The colon dataset uses whole-slide images as labeled source data and microscopy images as unlabeled target data for four-class polyp diagnosis.The classes are normal, adenoma, adenocarcinoma, and mucinous adenocarcinoma.
  • Dataset: Class imbalance is severe, and coarse annotation introduces noisy labels in source whole-slide images while target microscopy labels are clean.These conditions make the diagnosis task particularly difficult.
  • Results: CoUDA effectively adapts noisy source-domain knowledge to the unlabeled target domain in the colon diagnosis task.The authors describe the result as encouraging and relevant to medical diagnosis with limited labels.
  • Ablation Studies: Transferability-aware domain loss and the noise co-adaptation layer are relatively more important components in the Colon-A ablation study.The ablation results indicate that all components make empirical contributions.
  • Feature Visualization: CoUDA aligns colon-domain features well, whereas DANN, TCL, MCD, and CLAN fail to align them as effectively in the t-SNE visualization.The comparison uses adenoma features as an example of domain discrepancy and collaborative alignment.
  • Parameter Sensitivities: CoUDA is insensitive to α and η across {10^-3, 10^-2, 10^-1, 1} and performs best on Colon-A with α=0.1 and η=0.01.α adjusts domain loss and η controls diversity loss.

D. Application to General Images

Office-31 experiments test CoUDA on general images with controlled label noise, and the reported results support its effectiveness under domain shift and varying noise rates.

  • Dataset: Office-31 contains 4,652 images across 31 classes and three domains, producing six unsupervised transfer tasks.The domains are amazon, webcam, and dslr, with limited labeled source data under the standard setting.
  • Results: CoUDA handles domain discrepancy and label noise simultaneously better than the compared methods on Office-31.The authors also report that cooperation is helpful for the two challenges jointly.
  • Noise Transition Matrix: The noise co-adaptation layer approximates the true noise transition matrix within an acceptable error level on the W →D task.The comparison evaluates the estimated and true matrices under constructed label corruption.
  • Diverse Noise Rates: CoUDA consistently outperforms CLAN and TCL on Office-31 corrupted with noise rates of 0.2 and 0.4.Table IX reports the comparison using Macro F1 averaged across six transfer tasks.
  • Conclusion: The paper concludes that experiments on real-world medical and general images demonstrate CoUDA’s effectiveness and generalization.The authors expect applications beyond medicine, including crowdsourcing and financial domains.

I. DEFINITION AND PROOF FOR THEOREMS

The supplementary material defines the theoretical quantities used in CoUDA’s analysis and presents a generalization proof under explicit assumptions about losses, hypotheses, and label noise.

  • Definitions: The discrepancy distance measures the maximum loss difference between hypotheses across source and target distributions.It is defined over a hypothesis set H and loss function L.
  • Definitions: Rademacher complexity and its empirical counterpart quantify the hypothesis class’s sample-dependent complexity for the analysis.The definition uses independent uniform ±1-valued random variables.
  • Theorem 1: Theorem 1 bounds target classification performance using source classification loss, domain discrepancy, and the average loss between the best intra-class hypotheses.The bound assumes a symmetric loss satisfying the triangle inequality and an invertible, known noise transition matrix.
  • Proof: The proof transfers the single-peer generalization bound to the collaborative network because its prediction is a convex combination of the two peer predictions.The argument combines Proposition 1 with the two supporting lemmas.
  • Supporting Lemmas: Lemma Sup. 1 shows that domain discrepancy can be estimated from finite source and target samples using empirical Rademacher complexity.The result holds with the stated probabilistic guarantee over sampled data.
  • Supporting Lemmas: Lemma Sup. 2 shows that a classifier learned with noisy labels can estimate the optimal classifier learned with true labels under the stated noise assumptions.The lemma concerns modified focal loss for K-class classification.

C. Proof of Lemma Sup. 1

Lemma Sup. 1 bounds empirical domain discrepancy using Rademacher complexity, contraction, and triangle-inequality arguments. The proof also uses bounded domain logits and transferability-aware weights.

  • Rademacher-complexity bound: The proof starts from a Rademacher-complexity bound for function classes mapping inputs and labels to bounded losses.The bound holds with probability at least 1 − δ over sampled data and uniformly for all hypotheses.
  • Rademacher-complexity bound: The least-square domain loss is 2-Lipschitz, allowing the contraction lemma to control its complexity.The proposed domain loss relies on least-square, or L2 regression, loss.
  • Discrepancy decomposition: The proof bounds the Rademacher complexity of pairwise hypothesis differences through subadditivity and then combines empirical discrepancies using the triangle inequality.The final argument combines the relevant equations and adjusts the hypothesis class for the discriminator.
  • Bounded domain loss: Because discriminator logits lie in [0, 1] and transferability-aware weights lie in [1, 3], the domain loss is bounded by 3.This boundedness supports the subsequent high-probability generalization bound.

D. Proof of Lemma Sup. 2

Lemma Sup. 2 analyzes noisy-label learning by combining a noise co-adaptation assumption with a Rademacher-complexity generalization analysis. The proof reformulates focal loss as a modified logistic loss and derives a multi-class bound.

  • Noise adaptation: The analysis assumes the noise transition matrix Q is known and invertible, while the noise co-adaptation layer approximates Q.Under this setup, the noisy-label classifier is analyzed against the optimal classifier learned with true labels.
  • Noise adaptation: With sufficient noisy training data, the empirical optimum learned with noisy labels converges to the expected optimum associated with true labels.The proof states that the corresponding empirical and expected minimizers coincide in the limit under the stated assumptions.
  • Generalization analysis: The multi-class focal loss is reformulated as a single modified logistic loss involving noisy labels, prediction logits, and the class count K.The reformulation enables the subsequent complexity-based generalization analysis.
  • Generalization analysis: A concentration argument and a cited lemma yield a high-probability Rademacher-complexity upper bound on the multi-class generalization error.The bound holds with probability at least 1 − Kδ over m samples drawn from the noisy-label distribution.

II. ARCHITECTURE AND MORE DISCUSSIONS OF COUDA

CoUDA uses two peer networks combining a MobileNetV2-based backbone, noise co-adaptation layer, and discriminator. The discussions examine implementation choices for transferability, domain loss, and diversity loss.

  • Architecture: Each peer network contains a feature-extractor/classifier backbone, a noise co-adaptation layer, and a discriminator network.The two peer networks share the same architecture; the base network uses MobileNetV2 for resource and efficiency requirements.
  • Architecture: The noise co-adaptation layer uses K additional softmax layers to estimate transitions from each class to noisy pseudo-classes.These estimated transition probabilities produce predictions for noisy labels.
  • Architecture: The layer is initialized with zero weights and biases corresponding to a small uniform noise rate ϵ, whose value depends on the task.The indices i and j represent the true and noisy labels, respectively.
  • Architecture: For large class counts K, the K additional softmax layers create a scalability issue; a parameter-free alternative can reduce computational burden.The alternative is described as a suitable choice when K is very large.
  • Transferability and domain loss: Cosine distance is selected for transferability-aware weights because preliminary Colon-A experiments found no apparent differences among cosine, L1, and L2 distances, with cosine slightly better.The comparison is summarized in Table III.
  • Transferability and domain loss: Least-square domain loss is preferred because it is reported to improve domain confusion and stabilize training, and it outperforms GAN loss on Colon-A.The empirical comparison is reported in Table IV, whose displayed header includes accuracy, mean precision, mean recall, and macro F1.
  • Diversity loss: JS distance is used for diversity loss because preliminary experiments found no apparent differences among alternatives and relatively better performance for JS.The comparison is summarized in Table V.

D. Discussion about Best Intra-Class Hypotheses

The discussion examines the best intra-class hypothesis assumption and several design choices affecting CoUDA’s performance and cost. More peer networks improve performance but increase parameters and computation with diminishing gains.

  • Best Intra-Class Hypotheses: Domain adaptation assumes the average loss between the best intra-class hypotheses is small, because otherwise no hypothesis performs well on both domains.This condition is presented as necessary for conducting domain adaptation.
  • Best Intra-Class Hypotheses: Entropy loss was omitted because preliminary experiments found that it did not substantially improve performance.The authors retained the simpler proposed method instead.
  • Ensembling: Average prediction from two peer networks is used as the final prediction, while maximum ensembling performed slightly better in an alternative experiment.The average ensemble is described as effective in the main experiments.
  • Source annotation: Using more labeled source WSIs improves CoUDA’s performance, and the method performs well with only 20,000 labeled WSIs.This source-annotation analysis is conducted on Colon-A.
  • Peer-network count: Adding peer networks improves performance with gradually diminishing gains, while parameters, FLOPs, and memory consumption increase steadily.The trade-off is evaluated on Colon-A for different numbers of peer networks.
  • Training: Gradient Reverse Layer enables simultaneous end-to-end training of all network components, after backbone pretraining on source WSI data.CoUDA then trains with labeled WSIs and unlabeled MSIs.
Loading 2007.07222v1…