ARTIFICIAL INTELLIGENCE-BASED COMPUTATIONAL PATHOLOGY (AI-CPath) PLATFORM FOR LUNG CANCER

Abstract

Lung cancer stands as a malignant tumor which demonstrates both the highest frequency of occurrence and the highest death rate worldwide. Delayed diagnosis and scarce early detection strategies lead to this high incidence so researchers demand new diagnostic tools that would improve effectiveness and efficiency. Medical experts utilize histopathology as the current diagnostic standard for lung cancer yet slide examination by human pathologists takes considerable time along with showcasing inconsistent evaluation methods and demanding specialized training. Artificial intelligence (AI) combined with depth learning technology now provides a promising approach to improve diagnostic accuracy and decrease analyst workloads throughout pathology analysis. A digital pathology platform utilizes AI techniques for automating the classification procedure of lung cancer through histopathology image analysis. This platform utilizes Convolutional Neural Networks (CNNs) to distinguish between two NSCLC subtype groups which include adenocarcinoma and squamous cell carcinoma. The work investigates different CNN network designs to achieve maximum feature detection and classification results. One major obstacle in AI pathology emerges from insufficient high-quality annotated data that hampers generalization capabilities of models. The project addressed its issues by implementing methods to overcome the challenge of limited annotated medical data, contrastive learning (SimCLR) was integrated into the pipeline, enabling robust feature representation learning from unlabelled histopathology images. These self-supervised representations were then fine-tuned using a lightweight supervised classifier, demonstrating strong classification performance. This research delivers a reliable, accurate, and efficient AI-based classification tool that addresses the constraints of limited annotated data through contrastive learning, while also demonstrating the viability of both custom and pre-trained CNN models in digital pathology workflows for lung cancer diagnosis. Additionally, the careful composition of data augmentations was found to play a critical role in defining effective predictive tasks, especially within the contrastive learning framework, enabling the model to learn robust and transferable representations from unlabelled histopathological images.