Prediction Of Lung Metastasis For Osteosarcoma Within One Year After Surgery Based On A Deep Learning Model Integrating MRI And Digital Histopathology

BackgroundOsteosarcoma is the most common primary malignant bone tumor,which is highly malignant and prone to lung metastasis within a short period time,and 1-year lung metastasis is the main cause of death within osteosarcoma.How to accurately predict whether the patient will have lung metastasis during surgical treatment is the key factor to improve the diagnosis and treatment of osteosarcoma.Previous studies have shown that tumor metastasis can be predicted based on MRI radiomics and digital histopathology.However,the application of both in predicting lung metastasis of osteosarcoma is still in its infancy,and there is no research confirmation on the prediction of lung metastasis of osteosarcoma by combining the two to build a multigroup model.ObjectiveBased on the deep learning algorithm,we screened multiple biological indicators such as MRI radiomics and histopathological features related to lung metastasis of osteosarcoma,and constructed new multiple biomarkers to predict lung metastasis within one year of osteosarcoma.Method154 OS patients with preoperative MRI and postoperative pathologic specimens were enrolled from two hospitals and divided into a training(cohort D1)and an external test dataset(cohort D2).An automatic image analysis pipeline was constructed to extract quantitative features,and key features were selected to obtain a new signature for predicting 1-year LM,with the area under the curve(AUC)and C index as evaluation indicators.The one-year LM-free survival analysis(LMFS)and visualization analysis were further performed to compare the prognostic and clinical generalization abilities between image-based signatures and clinical variables.ResultsA new signature comprising MRI and HIS features was associated with LM.This signature has superior classification ability(AUC=0.804,C-index=0.75)to single MRI and HIS signatures on D2.Multivariate LMFS showed an HR of 3.31(p=1.44×10-2;95%CI,1.27-8.64)and 3.05(p=2.69 ×10-2;95%CI,1.148.20)for MRI-and HIS-based biomarkers,respectively,whereas the integrated signature demonstrated a more significant HR of 4.17(p=3.42× 10-3;95%CI,1.60-10.85).By contrast,clinical indicators showed HRs only ranging from 0.98 to 1.74.It demonstrated that the integrated signature could exceed the clinical indicators and provide more efficient LM clues in decision-making.ConclusionThe multimodal and multigroup model integrating MRI radiomics and histopathological features helps to more accurately predict the LM of osteosarcoma in one year,and determines a new integrated signature to effectively provide the LM and LMFS status of patients,which has greater guiding significance in achieving individualized precise treatment of osteosarcoma patients.