Artificial intelligence aids doctors in diagnosing necrotizing enterocolitis and predicting surgery using abdominal radiographs: a multicenter study

Comprehensive Summary

This retrospective study aimed to use convolutional neural networks (CNNs) to identify neonatal necrotizing enterocolitis (NEC) in preterm infants based on abdominal radiographs (AR). Retrospective data was included if the infant’s gestational age was less than 37 weeks, if they had complete clinical information, and no prior diagnosis of NEC/meeting specific diagnosis criteria. Abdominal radiographies (ARs) from time of diagnosis or from directly prior to surgery were chosen, and some were excluded depending on image quality or presence of congenital intestinal malformation; from this process, a total of 738 images, composed of 295 non-NEC, 307 medical-NEC, and 136 surgical-NEC cases, were chosen. All models were trained on ARs from the ImageNet dataset and were evaluated on accuracy, precision, recall, and F1 score. Data from subsets A, B, and a merged AB set were split into training (80%) and internal testing groups (20%). The best performing models during internal testing were then used in external testing using subset C data. This data was also separately evaluated twice by multiple clinicians of varying expertise: once with no external aid and again with an AI-made image heatmap of areas of interest. Not all doctors evaluated the images both times. The heatmaps showed overall high accuracy in pointing out significant bodily structures and improved prediction accuracy of clinicians, especially those with less experience. It should be noted that in set A, both the gestational age and birth weight in the non-NEC group were significantly higher than those in the NEC (medical + surgical) group (P<0.001). Additionally, the gestational age was significantly higher in the medical-NEC group compared to the surgical-NEC group (P<0.001). No such differences were observed in sets B and C, which had smaller sample sizes (P>0.05). The overall best models for sets A, B, and A+B were the Efficientnet-b3, b2, and b0 models, respectively. For external testing (set C), the EfficientNet-b0 model (trained on the A+B dataset) was the best overall, achieving high area under curve (AUC) for non-NEC and surgical-NEC data, while performing moderately for medical-NEC (AUC = 0.979, 0.996, and 0.924, respectively with values closer to one indicating more accuracy). EfficientNet-b0 significantly outperformed EfficientNet-b3 for non-NEC and surgical-NEC, and EfficientNet-b2 for non-NEC. The authors indicated limitations with data set diversity, inclusion of other types of infants, and exclusion of ARs with congenital intestinal malformation (which present similarly to NEC and could confuse the models).

Outcomes and Implications

NEC is one of the deadliest GI conditions in premature infants, and when caught early is managed with reduced intake, parenteral nutrition, and antibiotics. The condition is consistently monitored with ARs, though the interpretation of these images is subjective and cannot be used as a stand-alone method of diagnosis. Consequently, the use of CNNs and heatmap technology, such as Grad-CAM and deep feature factorization (DFF) in this study, may help junior clinicians increase prediction accuracy and diagnostic efficiency. It can also assist in better identification of severe NEC cases that may need surgery, lessening risk of fatality with added waiting time. With further testing, this CNN shows significant promise in helping pediatrics patients with NEC.