Hybrid Deep Learning Models for Enhanced Lung Cancer Detection from Medical Imaging

Abstract

Lung cancer is one of the most frequent causes of cancer related mortality in the world and early diagnosis is very essential in improving survival rates of the patients; nevertheless, traditional imaging technologies are usually characterized by poor results at differentiating benign and malign lesions, particularly in problematic cases such as small or ground-glass nodules, which leads to unnecessary high rates of false positive and false negative. To cope with them, this work proposes a new hybrid deep learning framework that combines convolutional neural networks (CNNs), long short-term memory (LSTM) recurring units, and transformer-based self-attention mechanisms to utilize both local spatial textures, temporal-sequential features, and contextual information of CT images. The multi-task learning implicit framework allows addressing issues of segmentation, detection, and classification in a single framework, enhancing both speed and quality.

On a large, strictly annotated dataset of 1,190 CT slices of 110 patients, the hybrid model secured the highest accuracy at 95 percent, significantly outperforming the best baselines, including CNN (90.5 percent), Transformer (91.2 percent), and LSTM-CNN (92.7 percent), as well as its precision and recall both to the tune of 95 percent. The gains in the model performance are supported by the fact that the F1 score is improved, and additional evidence of the model improvements is visualization tools such as attention maps and feature importance scores, which strengthen model interpretability and clinical trust.

These findings highlight the clinical opportunity of the hybrid method to enhance lung cancer diagnostic performance through increasing the accuracy of early detection, minimizing unnecessary invasive procedures, and providing personalized treatment plans. This paper presents the need to further substantiate in different multi-institutional data sets and stresses the future efforts on computational optimization and interpretability improvement, to convert the promising hybrid model into a scalable, clinically deployable decision support system to manage lung cancer.