A Deep Learning Based Suggested Model to Detect Necrotising Enterocolitis in Abdominal Radiography Images

Abstract

Despite decades of exploration into necrotising enterocolitis (NEC), we still lack the capacity to accurately diagnose the disease to improve outcomes in its management. Existing diagnostics struggle to delineate NEC from other neonatal intestinal diseases; it is also unable to highlight those likely to deteriorate to needing emergency life-saving surgery before it is too late. The diagnosis of NEC is heavily dependent on interpretation of radiological findings, especially abdominal radiography (AR) and abdominal ultrasound (AUS). Inter-expert variability in interpreting AR imaging, and in the case of AUS, performing and interpreting the test, remains an unresolved challenge. With the compounding impact of the shrinking radiology workforce, a novel approach is imperative. Computer assisted detection (CAD) and classification of abnormal pathology in medical imaging is a rapidly evolving field of clinical and biomedical research. This technology is widely used as a preliminary screening tool. This research paper proposes a deep learning-based model to classify AR images in an automated manner, generating class activation maps (CAM) from various imaging features consistent with NEC pathology, as agreed by expert consensus papers (in neonatology and paediatric radiology). It also compares it with conventional machine learning methods. The suggested model aims to produce heatmaps for various imaging features to highlight NEC pathology in AR (or in future AUS). Once the model is trained, validation is done through quantitative measures and visually by the attending radiologist (clinician) reviewing the validity of the colour maps highlighting the pathology of the AR image (future extension to AUS). As the volume of imaging data is increasing year by year, CAD can be a key strategy to assist radiology departments meet service needs. This technology can greatly assist in screening for NEC, improving the detection of NEC and potentially aid in the earlier identification of disease. Furthermore, it can fast track research cost effectively by creating big data through the automatic labeling of imaging data to create big-data for NEC databases.