Multimodal Medical Image Parsing with Task-Specific Low-Rank Adapters

Key points

• We study task-specific adaptation with Qwen3.5-9B for FLARE 2026 Task 3, covering six task families and seven imaging modalities.1
• We customize adapter ranks and prompts, rebalance rare classification targets, and supplement regression training with dermatology classification examples.1
• On the fixed internal holdout, the task-specific adapters achieve a dataset-macro balanced accuracy of 0.7791 for classification, a dataset-macro micro-F1 of 0.6286 for multi-label classification, a counting mean absolute error (MAE) of 11.73, a dataset-macro detection F1 of 0.8914, a dataset-macro regression MAE of 9.78, and a report-generation CRIMSON score of 0.8204.1
• The benefits vary across tasks, supporting the use of task-specific adapter configurations and image budgets within a shared medical vision-language model.1

Sources (1)

• [1]Multimodal Medical Image Parsing with Task-Specific Low-Rank Adapters
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 09:59 AM

  Adapting a single vision-language model to medical image parsing requires training choices that reflect the demands of each task.

  We study task-specific adaptation with Qwen3.5-9B for FLARE 2026 Task 3, covering six task families and seven imaging modalities.

Extractive summary: sentences quoted from the sources.