FusionClassNet: A Multi-Scale Feature Fusion Network with Contrastive Loss-Driven Classification for Enhanced Lung Tumor Image Representations

Abstract

Lung cancer remains a leading cause of mortality, making accurate subtype identification crucial for effective lung cancer diagnosis and treatment. Recent advancements in medical image classification, particularly through Convolutional Neural Networks (CNNs) and Transformers, have significantly improved the analysis of CT and histopathological images. However, CNNs are limited in capturing global features due to their restricted receptive fields. Alternatively, Transformers, while effective at modeling long-range dependencies, lack local inductive bias. To address these limitations, this paper introduces a novel multi-scale feature fusion network that integrates CNN and Transformer architectures. At its core is the Adaptive Feature Fusion (AFF) Block, which adaptively fuses hierarchical features from both networks, capitalizing on their strengths to improve classification performance. This module forms the feature extraction backbone and is complemented by a novel classification head. The classification head utilizes multi-head attention to perform cross-attention between class embeddings and image features. This process involves calculating cosine similarity for classification. The objective is to mitigate overfitting and address inter-class similarity issues. A contrastive loss function further maximizes the separation between class embeddings, enhancing classification accuracy. The proposed model was evaluated through multiple experiments on a clinical lung tumor surgical lesion slice image dataset. FusionClassNet achieved a classification accuracy of 88.68% and an F1-score of 83.28%. The experimental results demonstrate the effectiveness of the model in handling nonstandard medical images of lung tumors.