AI Trends in Medical Imaging: From Recognition to Workflow Automation (CVPR'26)

Simple Agents Outperform Experts in Biomedical Imaging Workflow Optimization

The paper proposes a lightweight AI agent that automatically generates pre‑processing and post‑processing code for existing biomedical image analysis tools, avoiding the need to train new models. Experiments on three real pipelines—Polaris (single‑molecule detection), Cellpose (cell instance segmentation), and MedSAM (medical image segmentation)—show that the agent consistently exceeds hand‑crafted expert optimizations, with especially large gains on MedSAM. Analysis of API and parameter spaces explains why the same agent behaves differently across tools, highlighting that simple, transparent agents can be sufficient for engineering‑heavy, low‑data workflow adaptation scenarios.

Paper URL: https://arxiv.org/pdf/2512.06006v1

DIQ: Difficulty‑Influence Quadrant for Efficient Medical Reasoning

DIQ selects a tiny subset of training data by jointly scoring each sample’s medical reasoning difficulty and its influence on model parameters. Using only 1 % of the data, DIQ matches or exceeds full‑data fine‑tuning on Huatuo and FineMed, and at 10 % data it outperforms random, perplexity‑based, similarity‑based, and LESS baselines. Human and LLM judges rate DIQ‑selected samples as more clinically plausible for diagnosis, safety checks, and evidence citation. Difficulty scores are derived from a BiomedBERT classifier, while influence is estimated via first‑order gradient dot products, making the method lightweight and reusable.

Paper URL: https://arxiv.org/pdf/2508.01450v3

CRAFT: Codebook‑Regulated Fine‑Tuning for Visual‑Language Models

CRAFT freezes the large language model and fine‑tunes only a discrete visual encoder anchored to a fixed codebook. Training combines a surrogate LLM alignment loss, a commitment loss, and a contrastive loss to ensure discrete tokens remain faithful to image content and understandable by the language model. During inference, a token‑rarity pruning step removes background or redundant tokens. Across ten benchmarks (IconQA, OCRVQA, ScienceQA, VQA‑RAD, EuroSAT, Flowers, Kvasir, PlantVillage, Cars, Dogs), CRAFT improves average accuracy by 13.51 % and reaches 68.58 % top‑1 accuracy in the strongest setting, outperforming LoRA, projector fine‑tuning, and continuous feature fine‑tuning.

Paper URL: https://arxiv.org/pdf/2602.19449v1

SPECTRE: Scaling Self‑Supervised and Cross‑Modal Pretraining for Volumetric CT Transformers

SPECTRE addresses 3‑D CT challenges—large token counts, anisotropic voxels, variable scan ranges, and noisy report supervision—by a two‑stage Transformer: a local ViT processes 3‑D windows, and a global ViT aggregates full scans. Self‑supervised learning and CT‑text cross‑modal alignment enable learning of both geometric structure and clinical semantics. Experiments show superior performance on tumor biomarker prediction, organ segmentation, and text‑to‑CT retrieval compared with most baselines, demonstrating that a purpose‑built pure‑Transformer architecture can capture both spatial detail and clinical meaning.

Paper URL: https://arxiv.org/pdf/2511.17209v2

Ultrasound‑CLIP: Semantic‑Aware Contrastive Pre‑training for Ultrasound Image‑Text Understanding

Ultrasound‑CLIP builds a 364 k image‑text pair dataset (US‑365K) covering 52 anatomical categories and defines a diagnostic taxonomy (UDT) with nine attributes. The model incorporates soft semantic labels and a heterogeneous graph encoder to capture relationships between lesions and attributes. Compared with generic CLIP and specialized baselines, Ultrasound‑CLIP achieves 59.61 % average classification accuracy (vs. 33.81 % for BiomedCLIP) and R@10 = 0.3745 for image‑to‑text retrieval, demonstrating that modality‑specific semantic modeling substantially improves ultrasound understanding.

Paper URL: https://arxiv.org/pdf/2604.01749v1

HyperST: Hierarchical Hyperbolic Learning for Spatial Transcriptomics Prediction

HyperST jointly models spot‑level and niche‑level features from H&E whole‑slide images and projects them into hyperbolic space. Hierarchical contrastive alignment and entailment constraints enable learning of "spot‑to‑niche" and "image‑to‑gene" relations. On four public datasets (HEST‑1K kidney, colorectal, skin, lung) HyperST improves PCC@200 by up to 10.95 % over the strongest baseline and attains AUROC = 0.719 for MSI‑H vs. MSS classification on TCGA‑COADREAD. Ablation studies confirm that removing the hierarchical hyperbolic alignment degrades performance, highlighting its importance.

Paper URL: https://arxiv.org/pdf/2511.22107

MoBind: Motion Binding for Fine‑Grained IMU‑Video Pose Alignment

MoBind aligns wearable IMU signals with video pose sequences via a hierarchical contrastive framework at token, body‑part, and global levels. Skeletons extracted from video reduce background interference, and each limb’s trajectory is matched to its corresponding IMU. On mRi, TotalCapture, and EgoHumans, MoBind reduces average temporal error to 0.05 s (TotalCapture) and 0.04 s (EgoHumans) and achieves 0.98–1.00 accuracy within a 200 ms tolerance, outperforming IMU2CLIP, DeSPITE, and SyncNet. Additional design choices—masked token prediction, multi‑sensor‑body part modeling, and hierarchical contrastive learning—address background noise, sensor‑body mapping, and fine‑grained synchronization challenges.

Paper URL: https://arxiv.org/pdf/2602.19004v1

SemVideo: Reconstructing Video from Brain Activity via Hierarchical Semantic Guidance

SemVideo reconstructs watched videos from fMRI signals by first extracting three‑level semantics from source videos using SemMiner: static anchor description, motion narrative, and global summary. A semantic‑aligned decoder (SAD) maps fMRI to these embeddings, while a motion‑adaptation decoder (MAD) models dynamic latent variables. On CC2017 and HCP‑7T datasets, SemVideo achieves top scores on 8 of 10 metrics, e.g., 2‑way‑V = 0.865, CLIP = 0.526, and EPE = 4.788, indicating superior semantic fidelity and temporal consistency. Ablations show that removing motion narration or MAD harms pixel‑level and temporal alignment, confirming the necessity of both semantic and motion components.

Paper URL: https://arxiv.org/pdf/2602.21819v2

GSR: Geometric‑Semantic Reasoner for Dual‑View X‑ray Prohibited Item Detection

GSR treats the second X‑ray view as a language‑like modality. Using Qwen3‑VL‑MoE‑8B, the model processes top‑view, side‑view, and textual queries in a structured chain‑of‑thought: first reasoning over the top view, then the side view, and finally producing a conclusion. The authors release DualXrayBench (45 613 dual‑view pairs, 12 prohibited‑item classes, 1 594 expert VQA samples). GSR‑8B attains 65.4 % accuracy, 70.6 % F1, and 52.3 % mIoU, outperforming GPT‑4o, Gemini‑2.5‑Pro, and single‑view baselines. Ablations demonstrate that merely adding a second view is insufficient; the structured reasoning over both views is essential for gains in geometric alignment and spatial relation understanding.

Paper URL: https://arxiv.org/pdf/2511.18385v1