ABSTRACT
The COVID-19 pandemic has highlighted the need for a tool to speed up triage in ultrasound scans and provide clinicians with fast access to relevant information. To this end, we propose a new unsupervised reinforcement learning (RL) framework with novel rewards to facilitate unsupervised learning by avoiding tedious and impractical manual labelling for summarizing ultrasound videos. The proposed framework is capable of delivering video summaries with classification labels and segmentations of key landmarks which enhances its utility as a triage tool in the emergency department (ED) and for use in telemedicine. Using an attention ensemble of encoders, the high dimensional image is projected into a low dimensional latent space in terms of: a) reduced distance with a normal or abnormal class (classifier encoder), b) following a topology of landmarks (segmentation encoder), and c) the distance or topology agnostic latent representation (autoencoders). The summarization network is implemented using a bi-directional long short term memory (Bi-LSTM) which utilizes the latent space representation from the encoder. Validation is performed on lung ultrasound (LUS), that typically represent potential use cases in telemedicine and ED triage acquired from different medical centers across geographies (India and Spain). The proposed approach trained and tested on 126 LUS videos showed high agreement with the ground truth with an average precision of over 80% and average F1 score of well over 44 ±1.7 %. The approach resulted in an average reduction in storage space of 77% which can ease bandwidth and storage requirements in telemedicine.
ABSTRACT
BACKGROUND AND OBJECTIVE: Lung ultrasound is an inherently user-dependent modality that could benefit from quantitative image analysis. In this pilot study we evaluate the use of computer-based pleural line (p-line) ultrasound features in comparison to traditional lung texture (TLT) features to test the hypothesis that p-line thickening and irregularity are highly suggestive of coronavirus disease 2019 (COVID-19) and can be used to improve the disease diagnosis on lung ultrasound. METHODS: Twenty lung ultrasound images, including normal and COVID-19 cases, were used for quantitative analysis. P-lines were detected by a semiautomated segmentation method. Seven quantitative features describing thickness, margin morphology, and echo intensity were extracted. TLT lines were outlined, and texture features based on run-length and gray-level co-occurrence matrix were extracted. The diagnostic performance of the 2 feature sets was measured and compared using receiver operating characteristics curve analysis. Observer agreements were evaluated by measuring interclass correlation coefficients (ICC) for each feature. RESULTS: Six of 7 p-line features showed a significant difference between normal and COVID-19 cases. Thickness of p-lines was larger in COVID-19 cases (6.27 ± 1.45 mm) compared to normal (1.00 ± 0.19 mm), P < 0.001. Among features describing p-line margin morphology, projected intensity deviation showed the largest difference between COVID-19 cases (4.08 ± 0.32) and normal (0.43 ± 0.06), P < 0.001. From the TLT line features, only 2 features, gray-level non-uniformity and run-length non-uniformity, showed a significant difference between normal cases (0.32 ± 0.06, 0.59 ± 0.06) and COVID-19 (0.22 ± 0.02, 0.39 ± 0.05), P = 0.04, respectively. All features together for p-line showed perfect sensitivity and specificity of 100; whereas, TLT features had a sensitivity of 90 and specificity of 70. Observer agreement for p-lines (ICC = 0.65-0.85) was higher than for TLT features (ICC = 0.42-0.72). CONCLUSION: P-line features characterize COVID-19 changes with high accuracy and outperform TLT features. Quantitative p-line features are promising diagnostic tools in the interpretation of lung ultrasound images in the context of COVID-19.