A Novel Artificial Intelligence System for the Prediction of Interstitial Lung Diseases

Title

A Novel Artificial Intelligence System for the Prediction of Interstitial Lung Diseases

Creator

Raju N.; Augustine D.P.; Chandra J.

Description

Interstitial lung disease (ILD) encompasses a spectrum of more than 200 fatal lung disorders affecting the interstitium, contributing to substantial mortality rates. The intricate process of diagnosing ILDs is compounded by their diverse symptomatology and resemblance to other pulmonary conditions. High-resolution computed tomography (HRCT) assumes the role of the primary diagnostic tool for ILD, playing a pivotal role in the medical landscape. In response, this study introduces a computational framework powered by artificial intelligence (AI) to support medical professionals in the identification and classification of ILD from HRCT images. Our dataset comprises 3045 HRCT images sourced from distinct patient cases. The proposed framework presents a novel approach to predicting ILD categories using a two-tier ensemble strategy that integrates outcomes from convolutional neural networks (CNNs), transfer learning, and machine learning (ML) models. This approach outperforms existing methods when evaluated on previously unseen data. Initially, ML models, including Logistic Regression, BayesNet, Stochastic Gradient Descent (SGD), RandomForest, and J48, are deployed to detect ILD based on statistical measures derived from HRCT images. Notably, the J48 model achieves a notable accuracy of 93.08%, with the diagnostic significance of diagonal-wise standard deviation emphasized through feature analysis. Further refinement is achieved through the application of Marker-controlled Watershed Transformation Segmentation and Morphological Masking techniques to HRCT images, elevating accuracy to 95.73% with the J48 model. The computational framework also embraces deep learning techniques, introducing three innovative CNN models that achieve test accuracies of 94.08%, 92.04%, and 93.72%. Additionally, we evaluate five full-training and transfer learning models (InceptionV3, VGG16, MobileNetV2, VGG19, and ResNet50), with the InceptionV3 model achieving peak accuracy at 78.41% for full training and 92.48% for transfer learning. In the concluding phase, a soft-voting ensemble mechanism amplifies training outcomes, yielding ensemble test accuracies of 76.56% for full-training models and 92.81% for transfer learning models. Notably, the ensemble comprising the three newly introduced CNN models attains the pinnacle of test accuracy at 97.42%. This research is poised to drive advancements in ILD diagnosis, presenting a resilient computational framework that enhances accuracy and ultimately betters patient outcomes within the medical domain. 2024, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.

Source

SN Computer Science, Vol-5, No. 1

Date

2024-01-01

Publisher

Springer

Subject

Deep learning; HRCT; ILD; Multi-label classification; Transfer learning

Coverage

Raju N., Department of Computer Science, CHRIST (Deemed to be University), Bengaluru, India; Augustine D.P., Department of Computer Science, CHRIST (Deemed to be University), Bengaluru, India; Chandra J., Department of Computer Science, CHRIST (Deemed to be University), Bengaluru, India

Rights

Restricted Access

Relation

ISSN: 2662995X

Format

Online

Language

English

Type

Article

Identifier

https://doi.org/10.1007/s42979-023-02524-3

https://www.scopus.com/inward/record.uri?eid=2-s2.0-85181497206&doi=10.1007%2fs42979-023-02524-3&partnerID=40&md5=a4f3647e56761e09098ce41e53c92520