Recursive Pixel Partitioning and Multi-Stream Deep Learning for Annotation-Efficient Lung Disease Classification

Abstract

Accurate lung disease classification from chest X-ray (CXR) images remains challenging in annotation-scarce clinical settings where pixel-level lung masks are unavailable. We propose a two-stage framework combining recursive pixel partitioning (RPP) with multi-stream deep learning for lung disease classification. Unlike anatomical segmentation, RPP hierarchically decomposes images into structured spatial sub-regions without manual masks, enabling multi-resolution feature extraction while preserving spatial organization. Dedicated streams process individual partitions, and their features are aggregated through collaborative learning to reduce information loss associated with single-stream convolutional neural networks. The framework is evaluated on two frontal CXR datasets: a balanced three-class dataset of 1,088 images (361 COVID-19, 365 Normal, and 362 Pneumonia) from the COVID-19 Radiography Database and Chest X-ray dataset, and the Shenzhen tuberculosis dataset (326 normal and 326 tuberculosis images). ImageNet-pretrained InceptionV3, MobileNetV2, Xception, ResNet101, and NASNetMobile are used for feature extraction. On the three-class task, the multi-stream model achieves 97.71% test accuracy, exceeding the best single-stream baseline (MobileNetV2, 94.50%) by 3.21 percentage points while converging approximately four times faster. On the Shenzhen dataset, it achieves 89.3% accuracy and 84.9% sensitivity, outperforming the reported segmentation-based pipelines on this split. Results are validated using stratified holdout and k-fold cross validation, supported by per-class metrics and an ablation study. The framework provides an annotation-efficient alternative for low resource settings without requiring radiologist-drawn lung masks. Received: 18 May 2026 | Revised: 24 August 2026 | Accepted: 16 October 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The dataset used in this study by the authors is publicly available and can be accessed at https://www.kaggle.com/datasets/nih-chest-xrays/data. Also, another data source used in this study is available at https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database.

Author Contribution Statement

Talib Iqball: Conceptualization, Methodology, Software, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Visualization. Umar Bashir: Validation, Writing – review & editing, Visualization. M. Arif Wani: Conceptualization, Methodology, Software, Resources, Writing – review & editing, Supervision, Project administration. Sheikh Umar Mushtaq: Validation, Formal analysis, Writing – original draft, Writing – review & editing, Visualization, Project administration.