Deep Looking in Digital Pathology: An Explainable Federated Learning Framework for Lung Cancer Classification

Background/Objectives: Lung cancer is the leading cause of cancer death worldwide, and growing slide volumes add to pathologists’ cognitive load. Deep learning can assist histopathological diagnosis, but two obstacles limit its clinical use: privacy regulations keep patient images in institutional “data silos”, and the “black-box” behavior of neural networks undermines clinical trust. Inspired by Topol’s Deep Medicine, we aimed to build a diagnostic platform that preserves privacy and explains its decisions. Methods: We developed the Lung Cancer Analysis System (LCAS), which combines federated learning (FL), explainable artificial intelligence (XAI), and a locally deployed medical large language model (LLM). Three public H&E datasets (LC25000, HistoSet-5×14, and CLID; 18,000 images; adenocarcinoma, squamous cell carcinoma, benign tissue) acted as three hospital nodes with source-induced non-independent and identically distributed (non-IID) data. DenseNet121, EfficientNet-B3, and MobileNetV3-Small were trained centrally and with FedAvg/FedProx for 20 rounds. Grad-CAM produced visual evidence, and MedGemma 1.5, served offline through LM Studio, converted predictions and heatmap descriptors into structured reports. Results: Centralized training reached 99.50% ± 0.21% accuracy with DenseNet121. Under non-IID federated training with FedProx, DenseNet121 and EfficientNet-B3 achieved 96.97% and 97.49% accuracy on an independent external blind test set of 958 images, whereas MobileNetV3-Small reached 91.23%. Accuracy remained 99.78% under Gaussian blur. t-SNE showed that global features clustered by tissue class rather than by hospital, and Grad-CAM highlighted glandular and tumor-nest regions consistent with histopathological criteria. Conclusions: FL, XAI, and on-premise LLM reporting can be integrated into a single workflow that keeps raw images within each institution while giving pathologists numerical, visual, and textual explanations, providing a practical path toward trustworthy cross-institutional computational pathology.