OncoSight – AI Powered Early Cancer Detection System

Abstract: Cancer remains one of the foremost causes of mortality worldwide, with early and accurate diagnosis being the most decisive factor in improving patient survival outcomes. Traditional cancer detection methods rely on manual examination of histopathological tissue slides and CT scan images by radiologists and pathologists — a process that is time-consuming, resource-intensive, and inherently prone to human error. OncoSight is a web-based AI diagnostic portal developed using the Django framework to address this critical gap. The system integrates two deep learning models into a unified, browser-accessible interface: a ResNet50 model trained on the LC25000 histopathological dataset for three- class lung tissue classification (Adenocarcinoma, Squamous Cell Carcinoma, Normal), and a weighted ensemble of EfficientNetB3 (25%), EfficientNetB4 (45%), and ResNet50V2 (30%) trained on the LIDC-IDRI CT scan dataset (1,010 patients) for binary lung nodule malignancy detection using focal loss and AdamW optimiser with a malignancy threshold of 0.40. Grad-CAM heatmap visualisations are generated for every prediction, highlighting regions that most influenced the AI decision. The system achieves a weighted ensemble ROC-AUC of 0.8776, delivers predictions within seconds, and provides dedicated per-patient diagnostic report pages. OncoSight addresses the real-world gap of no accessible, clinically deployable AI cancer detection platform while being architecturally designed for future multi-cancer expansion.

Keywords: Cancer Detection, Deep Learning, ResNet50, EfficientNet, Transfer Learning, Grad-CAM, LIDC-IDRI, Histopathological Classification, Ensemble Learning, Django.