A Swin Transformer based on multi-directional-shift window attention and inductive bias for diagnosis of pleural effusion

In the field of healthcare, deep learning has shown promise in addressing diagnostic challenges. However, existing methods often struggle with generalization due to overfitting on non-discriminative features and limited datasets. To address these limitations, Ultra-Multi-SWIN is introduced as a novel deep learning model for pleural effusion diagnosis using ultrasound images. The model incorporates physician-inspired inductive biases into its architecture, enabling it to focus on discriminative features while avoiding overfitting to irrelevant information. Specifically, a multi-directional-shift window structure captures spatial features dependent on direction, and a MASK-based masking module suppresses redundant non-ultrasound features. A dataset comprising 50 subjects and four levels of pleural effusion severity (large, moderate, small, none) is established to evaluate the model's performance. Experimental results demonstrate that Ultra-Multi-SWIN achieves state-of-the-art performance, with average accuracies of 0.988 (subject-dependent) and 0.952 (subject-independent). Visualization and ablation studies further confirm the model's ability to generalize effectively by focusing on clinically relevant regions. The open-source code is released at Ultra-Multi-SWIN, promoting broader adoption and future research.