Video-Based CSwin Transformer Using Selective Filtering Technique for Interstitial Syndrome Detection

Interstitial lung diseases (ILD) significantly impact health and mortality, affecting millions of individuals worldwide. During the COVID-19 pandemic, lung ultrasonography (LUS) became an indispensable diagnostic and management tool for lung disorders. However, utilising LUS to diagnose ILD requires significant expertise. This research aims to develop an automated and efficient approach for diagnosing ILD from LUS videos using AI to support clinicians in their diagnostic procedures. We developed a binary classifier based on a state-of-the-art CSwin Transformer to discriminate between LUS videos from healthy and non-healthy patients. We used a multi-centric dataset from the Royal Melbourne Hospital (Australia) and the ULTRa Lab at the University of Trento (Italy), comprising 60 LUS videos. Each video corresponds to a single patient, comprising 30 healthy individuals and 30 patients with ILD, with frame counts ranging from 96 to 300 per video. Each video is annotated using the corresponding medical report as ground truth. The datasets used for training the model underwent selective frame filtering, including reduction in frame numbers to eliminate potentially misleading frames in non-healthy videos. This step was crucial because some ILD videos included segments of normal frames, which could be mixed with the pathological features and mislead the model. To address this, we eliminated frames with a healthy appearance, such as frames without B-lines, thereby ensuring that training focused on diagnostically relevant features. The trained model was assessed on an unseen, separate dataset of 12 videos (3 healthy and 9 ILD) with frame counts ranging from 96 to 300 per video. The model achieved an average classification accuracy of 91%, calculated as the mean of three testing methods: Random Sampling (92%), Key Featuring (92%), and Chunk Averaging (89%). In RS, 32 frames were randomly selected from each of the 12 videos, resulting in a classification with 92% accuracy, with specificity, precision, recall, and F1-score of 100%, 100%, 90%, and 95%, respectively. Similarly, KF, which involved manually selecting 32 key frames based on representative frames from each of the 12 videos, achieved 92% accuracy with a specificity, precision, recall, and F1-score of 100%, 100%, 90%, and 95%, respectively. In contrast, the CA method, where the 12 videos were divided into video segments (chunks) of 32 consecutive frames, with 82 video segments, achieved an 89% classification accuracy (73 out of 82 video segments). Among the 9 misclassified segments in the CA method, 6 were false positives and 3 were false negatives, corresponding to an 11% misclassification rate. The accuracy differences observed between the three training scenarios were confirmed to be statistically significant via inferential analysis. A one-way ANOVA conducted on the 10-fold cross-validation accuracies yielded a large F-statistic of 2135.67 and a small p-value of 6.7 × 10−26, indicating highly significant differences in model performance. The proposed approach is a valid solution for fully automating LUS disease detection, aligning with clinical diagnostic practices that integrate dynamic LUS videos. In conclusion, introducing the selective frame filtering technique to refine the dataset training reduced the effort required for labelling.