# Lung Nodule Malignancy Classification Integrating Deep and Radiomic Features in a Three-Way Attention-Based Fusion Module

## TL;DR

This paper introduces a new AI framework that combines deep learning and radiomic features to classify lung nodules as either invasive or non-invasive with high accuracy.

## Contribution

A novel three-way attention-based fusion module (I-VISTA) that integrates deep learning and radiomic features for improved lung nodule classification.

## Key findings

- The I-VISTA framework achieved 93.93% overall accuracy in classifying subsolid lung nodules.
- The hybrid model outperformed standalone deep learning and radiomic models in differentiating nodule groups.
- The framework demonstrated high sensitivity (92.66%) and specificity (94.99%) with an AUC of 0.93.

## Abstract

In this study, we propose a novel hybrid framework for assessing the invasiveness of an in-house dataset of 114 pathologically proven lung adenocarcinomas presenting as subsolid nodules on Computed Tomography (CT). Nodules were classified into group 1 (G1), which included atypical adenomatous hyperplasia, adenocarcinoma in situ, and minimally invasive adenocarcinomas, and group 2 (G2), which included invasive adenocarcinomas. Our approach includes a three-way Integration of Visual, Spatial, and Temporal features with Attention, referred to as I-VISTA, obtained from three processing algorithms designed based on Deep Learning (DL) and radiomic models, leading to a more comprehensive analysis of nodule variations. The aforementioned processing algorithms are arranged in the following three parallel paths: (i) The Shifted Window (SWin) Transformer path, which is a hierarchical vision Transformer that extracts nodules’ related spatial features; (ii) The Convolutional Auto-Encoder (CAE) Transformer path, which captures informative features related to inter-slice relations via a modified Transformer encoder architecture; and (iii) a 3D Radiomic-based path that collects quantitative features based on texture analysis of each nodule. Extracted feature sets are then passed through the Criss-Cross attention fusion module to discover the most informative feature patterns and classify nodules type. The experiments were evaluated based on a ten-fold cross-validation scheme. I-VISTA framework achieved the best performance of overall accuracy, sensitivity, and specificity (mean ± std) of 93.93 ± 6.80%, 92.66 ± 9.04%, and 94.99 ± 7.63% with an Area under the ROC Curve (AUC) of 0.93 ± 0.08 for lung nodule classification among ten folds. The hybrid framework integrating DL and hand-crafted 3D Radiomic model outperformed the standalone DL and hand-crafted 3D Radiomic model in differentiating G1 from G2 subsolid nodules identified on CT.