A Meta-learning-based Robust Federated Learning for Diagnosing Lung Adenocarcinoma and Tuberculosis Granulomas

Chen Yuyao¹Lei Liu²Bao Feng²Yehang Chen²Jun Xu²Huan Lin³Kunwei Li⁴Xiaodong Chen⁵Yuting Ke⁵Haoyang Zhou²Qinghui Hu²Qinggeng Jin⁶Wansheng Long⁷Qiong LI^8*Xiangmeng Chen^1*

• ¹Nanxishan Hospital of Guangxi Zhuang Autonomous Region, Guilin, China
• ²Guilin University of Aerospace Technology, Guilin, China
• ³Guangdong Provincial People's Hospital, Guangzhou, China
• ⁴Fifth Affiliated Hospital of Sun Yat-sen University Department of Radiology, Zhuhai, China
• ⁵Affiliated Hospital of Guangdong Medical University Department of Radiology, Zhanjiang, China
• ⁶Guangxi University School of Electrical Engineering, Nanning, China
• ⁷Jiangmen Central Hospital, Jiangmen, China
• ⁸Sun Yat-Sen University Cancer Center Department of Radiotherapy, Guangzhou, China

Background: Differentiating between lung adenocarcinoma (LAC) and tuberculosis granuloma (TBG) of solitary pulmonary solid nodules (SPSNs) based on CT images alone is a daunting task for clinical diagnosis. Thus, it is crucial to fully utilize CT imaging data to explore effective noninvasive diagnostic methods to improve the identification of TBG and LAC. Purpose: This study aimed to leverage CT imaging datasets from multiple hospitals for the diagnosis of TBG and LAC in SPSNs. It achieved this by deploying a meta-learning method within a federated learning framework while protecting data privacy. Methods: A total of 1,026 patients, along with their CT images of solitary pulmonary solid nodules (SPSNs) and corresponding clinical data, were collected from six medical institutions. Subsequently, the data from these six institutions were systematically partitioned into five cohorts. Each cohort was divided into two parts: the training set and the test set. A meta-learning-based robust federated learning model by training set data was proposed to construct personalized federated learning signatures (PFLS) without uploading raw data from each medical institutions. Receiver operating characteristic curve (ROC), area under curve (AUC), decision curve analysis (DCA), net reclassification improvement (NRI) and integrated discrimination improvement (IDI) are used to analyze the performance of the PFLS. Results: The PFLS trained by the proposed meta-learning-based robust federated learning framework shows superior performance compared to alternative methods. The AUC range on the training sets of the five cohorts is 0.866-0.939, AUC range on the testing sets is 0.808-0.927). The significant difference of AUC between the proposed method and the clinical model was demonstrated by the NRI and IDI. The decision curves indicated a higher net benefit of our proposed method. Conclusion: The PFLS mitigates overfitting issues arising from limited sample size in local hospitals. It also alleviates the problem that a single global model is not applicable to all hospitals due to the heterogeneity of data distribution among different hospitals.