A combi-elasto ultrasound based deep learning model on liver fibrosis staging for suspected MASLD/MASH patients

Metabolic dysfunction-associated steatotic liver disease (MASLD) and metabolic dysfunction-associated steatohepatitis (MASH) weakened the diagnostic ability of traditional elastography ultrasound on liver fibrosis. Combi-elasto ultrasound can simultaneously reflect the information of fibrosis degree and fat accumulation, thus have better potential and generality for liver fibrosis. We aim to develop a combi-elasto ultrasound based deep learning (DL) model for fibrosis staging in suspected MASLD/MASH patients. Patients with overweight, type-2 diabetes, hypertension or dyslipidemia, and without alcoholic, drug-induced, and genetic hepatitis history were defined as suspected MASLD/MASH patients. Fibrosis stages were classified as none or mild (F0/F1), moderate (F2), and severe (F3/F4). Combi-elasto images from ten centers were collected to build the DL model. OpenCV was applied to automatically extract image information, Resnet-50 was used to analyze image feature, and adversarial training (AT) algorithm was introduced to enhance model generality. Traditional DL model without AT algorithm, and other seven classical fibrosis models were also built to compare with AT-DL model. Subgroup analysis on pathological confirmed normal, MASLD and MASH patients were performed to test model generality. From January 2022 to December 2023, 295 patients from eight centers were divided into training (n = 204) and validation (n = 91) sets. 101 patients from two independent centers were collected as external test set. In test set, AT-DL model achieved significantly higher accuracy than traditional DL model (0.871 vs. 0.812, p = 0.021) and all other classical models (0.545–0.762, all p < 0.05). For pathological confirmed normal patients, AT-DL and traditional DL models had comparable accuracy (0.881 vs. 0.905, p = 0.274). However, for pathological confirmed MASLD and MASH patients, AT-DL model showed significantly accuracy than traditional DL model (0.900 vs. 0.800, p = 0.028 and 0.828 vs. 0.690, p = 0.003). Combi-elasto ultrasound based AT-DL model had satisfactory performance on fibrosis staging in suspected MASLD/MASH patients, and could reliable non-invasive monitoring of fibrosis progression. Combi-elasto based deep learning mode had satisfactory performance on fibrosis staging in suspected metabolic dysfunction-associated steatotic liver disease and metabolic dysfunction-associated steatohepatitis patients. Combi-elasto based adversarial training deep learning model achieved satisfactory accuracy on fibrosis staging (87.1%, 95% CI: 81.2–92.1%) in suspected MASLD/MASH patients. Compared with traditional deep learning model, adversarial training algorithm significantly improved accuracy in pathological confirmed MASLD (from 80.0% to 90.0%, p = 0.012) and MASH patients (from 69.0% to 82.8%, p = 0.003).