Explainable Split-Learning-Based Framework for Accurate Pulmonary Nodule Classification

Abstract

Lung cancer rates are the highest among cancers, making it the leading cause of death worldwide. With advances in new technologies and diverse diagnostic methods, Computer-Aided Diagnosis Systems (CADx) have improved pulmonary nodule classification with notable accuracy and speed. However, limited data availability and privacy concerns remain significant challenges, in addition to the reported rates of false negatives and false positives. This work aims to develop an approach based on collaborative feature extraction between multiple centers, thus achieving data efficiency and diversity while ensuring privacy and reducing false positives and false negatives. This work proposes a new explainable feature-based split learning approach using diverse Computed Tomography (CT) scan datasets to evaluate data diversity and privacy. It adopts a split ResNet-50 architecture on the client side for feature extraction. On the server side, a hybrid 2D-CNN combined with an attention mechanism is used for final classification and decision-making. The architecture was evaluated using two ablation studies based on ConvNeXt-Tiny and EfficientNetB0. In addition, the model was tested on two external datasets to assess its robustness and generalizability, and with Local Interpretable Model-agnostic Explanations (LIMEs) and Grad-CAM to assess trustworthiness. This proposed approach showed an accuracy and F1-score of 99.38%, with a 1.23% false negative rate and zero false positives. Moreover, when tested on totally unseen datasets, the approach achieved an accuracy and an F1-score of 99.28% on the first dataset, with 1.24% false negatives and 0% false positives. In addition, when tested on the second dataset, the results indicate an ability to generalize, with 95.74% accuracy, with false negative and false positive rates of 7.07% and 1.41%, respectively.

keywords: CT scan, diagnosis, explainability, lung cancer, ResNet-50, split learning, XAI