Related Experiment Video
Updated: Jun 28, 2025
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
6.8K
An investigation into augmentation and preprocessing for optimising X-ray classification in limited datasets: a case
Franciszek Nowak1, Ka-Wai Yung2, Jayaram Sivaraj3
1Wellcome/EPSRC Centre for Interventional and Surgical Sciences, Department of Medical Physics and Biomedical Engineering, UCL, London, UK. franciszek.nowak.23@ucl.ac.uk.
Summary
Image preprocessing and augmentation significantly improve deep learning models for diagnosing necrotising enterocolitis (NEC) in premature infants using limited abdominal X-ray data.
Area of Science:
- Medical Imaging Analysis
- Artificial Intelligence in Healthcare
- Neonatal Medicine
Background:
- Deep learning for medical image analysis requires large datasets, which are difficult to obtain for rare diseases like necrotising enterocolitis (NEC).
- Necrotising enterocolitis (NEC) is a critical condition in premature neonates with challenging radiological diagnosis.
- Data augmentation and preprocessing are essential strategies to overcome data scarcity in medical AI development.
Purpose of the Study:
- To investigate the impact of various image augmentation and preprocessing techniques on the performance of computer-aided diagnosis (CAD) models for NEC.
- To propose optimized preprocessing pipelines (Pr-1 and Pr-2) for enhancing the detection of subtle NEC findings in abdominal X-rays (AXRs).
- To develop reliable CAD models for a challenging three-class NEC classification task using a limited dataset.
Main Methods:
- Utilized a dataset of 1090 AXRs from 364 patients with NEC.
- Evaluated geometric augmentations (Translation, Flipping, Occlusion) and color augmentations (Equalisation) using a ResNet-50 backbone.
- Introduced two novel preprocessing pipelines (Pr-1, Pr-2) focusing on color contrast and edge enhancement.
Main Results:
- Geometric augmentations, particularly Translation (+6.2%), improved model performance; Flipping and Occlusion showed negative impacts.
- Color augmentations provided modest improvements; Equalisation yielded slight gains.
- The proposed Pr-1 and Pr-2 pipelines increased model accuracy by +2.4% and +1.7%, respectively.
- Combining optimized pipelines with geometric augmentation achieved a maximum performance increase of 7.1% for robust NEC classification.
Conclusions:
- Image preprocessing demonstrates significant, previously unreported potential for AXR classification tasks with limited data.
- The developed techniques and findings offer a benchmark for automated NEC detection and classification from AXRs.
- These methods are extensible to other medical imaging tasks requiring reliable classifier models with scarce datasets.
