Frequency-Aware Multi-Head Learning for Long-Tailed Medical Concept Detection in ImageCLEFmed Caption 2026

Frequency-Aware Multi-Head Learning for Long-Tailed Medical Concept Detection in ImageCLEFmed Caption 2026

Summary

This paper describes the submission of team basharderar to the ImageCLEFmed Caption 2026 Concept Detection Task. The task aims to predict clinically relevant Unified Medical Language System (UMLS) concepts from radiology images in a large-scale, sparse, and hi...

This paper describes the submission of team basharderar to the ImageCLEFmed Caption 2026 Concept Detection Task. The task aims to predict clinically relevant Unified Medical Language System (UMLS) concepts from radiology images in a large-scale, sparse, and highly imbalanced multi-label setting. We explore a frequencyaware multi-head learning strategy built on a Swin Transformer visual backbone. The concept vocabulary is divided into frequency-based tiers, and each tier is assigned a dedicated prediction head. This design is used together with Asymmetric Loss, frequency-group loss weighting, exponential moving average, multilabel stratified validation, and post-hoc threshold calibration. To address long-tailed evaluation more carefully, we report additional ablation and frequency-tier analyses comparing the proposed model with two single-head Swin-Base baselines: a standard Asymmetric Loss baseline and a class-weighted BCE baseline with post-hoc thresholds. The proposed frequency-aware model achieved a three-fold validation Micro-F1 of 0.5421 under per-head threshold calibration, which is comparable to the singlehead ASL baseline and higher than the class-weighted single-head baseline under the global Micro-F1-oriented operating point. The frequency-tier analysis shows that aggregate Micro-F1 is strongly dominated by common concepts, while rare-concept recovery depends heavily on the selected operating point. On the official hidden test set, team basharderar achieved a primary score of 0.5725 and a secondary score of 0.9419, ranking 4th in the official task leaderboard. Overall, our results highlight the trade-off between aggregate Micro-F1 and rare-concept recovery in long-tailed medical concept detection, and suggest that frequency-aware calibration is an important component for analyzing such systems.

Keywords

medical concept detection ImageCLEFmed Caption 2026 multi-label classification long-tailed learning frequencyaware learning Swin Transformer threshold calibration

Researchers

Publication Files