G.TAN1, A.LIU1, C.M.NYEIN1, T.DU2, C.U.UBEYNARAYANA1
Khoo Teck Puat Hospital1, Nanyang Polytechnic2
Fluid overload is a significant problem among dialysis patients, with pulmonary oedema being the most serious consequence. B lines on lung ultrasound(LUS) is a reliable method of estimating lung water with higher sensitivity compared to clinical examination. The widespread use of LUS in dialysis patients is limited by a shortage of trained staff in interpreting LUS images.
Our study aims to use artificial intelligence(AI) to interpret LUS B-lines to measure lung water in dialysis patients.
LUS is performed by 2 nephrologist and a dialysis nurse. B-line count is interpreted by the nephrologists and nurse, and compared to B-line count using AI. Our primary outcome is the correlation between automated and physician interpretation of LUS images in detecting lung water in dialysis patients.
A Mask-RCNN deep learning network is designed and developed to detect B-lines in LUS. The network is trained using labelled data through transfer learning to recognize and segment B-lines from other fresh LUS images. After that, a B-line tracking algorithm is developed to count the number of B-lines.
LUS performed on 30 dialysis patients showed statistically significant correlation between B line count interpreted by trained healthcare personnel compared to AI (p <0.001).
Our study shows that automated LUS has good correlation with interpretation done by trained staff. Automated LUS will allow more widespread use of LUS in dialysis patients to detect early fluid overload in dialysis patients and prevent complications of pulmonary oedema and hospital admissions.