Label Transfer Hypothesis: A Clinical Prior Knowledge-Guided Approach for Disease Diagnosis

ICLR 2026 Conference Submission19022 Authors

Keywords: Label Transfer, Medical Image Analysis, Disease Diagnosis

TL;DR: We propose Label Transfer Hypothesis: using lesion segmentation as implicit diagnostic labels to enable disease classification without requiring extensive expert annotations.

Abstract:

Deep learning in medical image analysis requires large-scale, high-quality annotated datasets that are expensive and time-consuming to obtain due to extensive expert involvement. Most existing approaches rely on supervised learning, severely limiting practical deployment given annotation scarcity.

To address this limitation, we propose the Label Transfer Hypothesis (LTH), a theoretical framework for tackling annotation scarcity. The core hypothesis holds that when diseases present characteristic pathological features, precise lesion segmentation guided by clinical diagnosis and treatment logic can act as "implicit diagnostic labels" for disease classification—this enables knowledge transfer from segmentation to classification tasks. This approach not only reduces annotation requirements while retaining the advantages of supervised classification, but also leverages the combination of the label transfer method and clinical diagnosis and treatment logic to obtain more reliable diagnoses.

We validate LTH on diabetic macular edema (DME) and retinal vein occlusion (RVO) classification tasks. Results demonstrate that LTH achieves performance comparable or superior to supervised methods while requiring significantly fewer labeled data.The code will be released after acceptance.

This work contributes: (1) A pioneering theoretical framework bridging segmentation and classification through clinical knowledge integration; (2) Demonstrated feasible knowledge transfer maintaining competitive performance with reduced supervision; (3) A scalable solution for resource-constrained healthcare settings, particularly beneficial for medically underserved regions.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 19022