Bowen Wang

Progressive Fibrosing Interstitial Lung Disease (PF-ILD) is a severe phenotype of Interstitial Lung Disease (ILD) with a poor prognosis, typically requiring prolonged clinical observation and multiple CT examinations for diagnosis. Such requirements delay early detection and treatment initiation. To enable earlier identification of PF-ILD, we propose ILD-Slider, a parameter-efficient and lightweight deep learning framework that enables accurate PF-ILD identification from a limited number of CT slices. ILD-Slider introduces anatomy-based position markers (PMs) to guide the selection of representative slices (RSs). A PM extractor, trained via a multi-class classification model, achieves high PM detection accuracy despite severe class imbalance by leveraging a peak slice mining (PSM)-based strategy. Using the PM extractor, we automatically select three, five, or nine RSs per case, substantially reducing computational cost while maintaining diagnostic accuracy. The selected RSs are then processed by a slice-level 3D Adapter (Slider) for PF-ILD identification. Experiments on 613 cases from The University of Osaka Hospital (UOH) and the National Hospital Organization Osaka Toneyama Medical Center (OTMC) demonstrate the effectiveness of ILD-Slider, achieving an AUPRC of 0.790 (AUROC 0.847) using only five automatically extracted RSs. ILD-Slider further validates the feasibility of diagnosing PF-ILD from non-contiguous slices, which is particularly valuable for real-world and public datasets where contiguous volumes are often unavailable. These results highlight ILD-Slider as a practical and efficient solution for early PF-ILD identification.

Large-scale models trained on extensive datasets have become the standard due to their strong generalizability across diverse tasks. In-context learning (ICL), widely used in natural language processing, leverages these models by providing task-specific prompts without modifying their parameters. This paradigm is increasingly being adapted for computer vision, where models receive an input–output image pair, known as an in-context pair, alongside a query image to illustrate the desired output. However, the success of visual ICL largely hinges on the quality of these prompts. To address this, we propose Enhanced Instruct Me More (E-InMeMo), a novel approach that incorporates learnable perturbations into in-context pairs to optimize prompting. Through extensive experiments on standard vision tasks, E-InMeMo demonstrates superior performance over existing state-of-the-art methods. Notably, it improves mIoU scores by 7.99 for foreground segmentation and by 17.04 for single object detection when compared to the baseline without learnable prompts. These results highlight E-InMeMo as a lightweight yet effective strategy for enhancing visual ICL.