# Design of an integrated model using U-Net, DeepSurv, and cross-attention for lung cancer classification and survival prediction

## TL;DR

This paper introduces a deep learning framework combining U-Net, DeepSurv, and cross-attention to improve lung cancer classification and survival prediction using multimodal data.

## Contribution

The framework achieves high tumor segmentation accuracy and robust survival predictions with adaptive learning for real-time clinical use.

## Key findings

- Tumor segmentation accuracy ranges from 90 to 95% Dice similarity.
- Lung cancer subtype classification accuracy improves between 85% and 90%.
- Survival rate predictions achieve a C Index of ~0.75–0.80.

## Abstract

Lung cancer ranks within the highest mortality rates among cancerous diseases; hence, its detailed classification and survival rate prediction are of utmost importance. Most existing approaches for the classification and prognosis prediction in lung cancer share a critical deficiency: they are either single-modality or fail to learn complex, nonlinear interactions between distinct data types. However, none of these traditional models iteratively refines segmentation with the requisite accuracy to embed continuous flow of new patient data without degradation in performance. We hence propose an Iterative Multi-Model Deep Learning Framework for improved classification of lung cancer subtypes and predictions of survival rates. Our proposed work uses the U-Net model, which refines features extracted iteratively to improve precision in segmented regions. For lung cancer subtypes classification, feature-level fusion is done by using CNN for spatial features extracted from both radiological and histopathology images and using an MLP for genomic data samples. The DeepSurv model extends the Cox proportional hazards model with deep learning to handle complex, multi-dimensional clinical, imaging, and genomic data for survival rate prediction. Bayesian optimization is used to optimize the hyperparameter tuning process, whereas EWC empowers this approach with real-time survival predictions, thus enabling incremental learning without catastrophic forgetting. This is further reinforced by a multimodal attention mechanism that ensures the most discriminative features from each modality are taken into consideration by the model. The contributions of this work consist of an improvement in tumor segmentation accuracy with results that range from 90 to 95% Dice similarity, a raise of accuracy in lung cancer subtype classification between 85% and 90%, and robust survival rate predictions with a C Index of ~ 0.75–0.80. Besides, our adaptive learning approach can continuously improve our model to make it fit for real-time clinical applications. The framework will present an end-to-end solution for the diagnosis and prognosis of lung cancer.