Generalisability of fetal ultrasound deep learning models to low-resource imaging settings in five African countries

Abstract

Most artificial intelligence (AI) research and innovations have concentrated in high-income countries, where imaging data, IT infrastructures and clinical expertise are plentiful. However, slower progress has been made in limited-resource environments where medical imaging is needed. For example, in Sub-Saharan Africa, the rate of perinatal mortality is very high due to limited access to antenatal screening. In these countries, AI models could be implemented to help clinicians acquire fetal ultrasound planes for the diagnosis of fetal abnormalities. So far, deep learning models have been proposed to identify standard fetal planes, but there is no evidence of their ability to generalise in centres with low resources, i.e. with limited access to high-end ultrasound equipment and ultrasound data. This work investigates for the first time different strategies to reduce the domain-shift effect arising from a fetal plane classification model trained on one clinical centre with high-resource settings and transferred to a new centre with low-resource settings. To that end, a classifier trained with 1792 patients from Spain is first evaluated on a new centre in Denmark in optimal conditions with 1008 patients and is later optimised to reach the same performance in five African centres (Egypt, Algeria, Uganda, Ghana and Malawi) with 25 patients each. The results show that a transfer learning approach for domain adaptation can be a solution to integrate small-size African samples with existing large-scale databases in developed countries. In particular, the model can be re-aligned and optimised to boost the performance on African populations by increasing the recall to [Formula: see text] and at the same time maintaining a high precision across centres. This framework shows promise for building new AI models generalisable across clinical centres with limited data acquired in challenging and heterogeneous conditions and calls for further research to develop new solutions for the usability of AI in countries with fewer resources and, consequently, in higher need of clinical support.

Figure 1

Image examples of the maternal-fetal US categories from our multi-centre dataset.

Figure 2

Generalisation performance of a model trained with different number of samples from the Spanish dataset evaluated in the same clinical centre, in a new centre in Europe (Denmark) and five datasets from Africa (Malawi, Egypt, Uganda, Ghana, Algeria).

Figure 3

The AUC score as a result of combining the dataset acquired in Spain with the US samples acquired in each African centre independently or after fine-tuning the Spanish model, trained with varying sample size, with US images from each African centre independently.

Figure 4

The evaluation of the model after being trained in Spain with all available samples and fine-tuned using a different number of patients from each African centre, as compared to the performance of the model directly trained with the African US images.

Figure 5

Results on common plane classification with and without transfer learning using 12 patients of each African centre in Malawi and Algeria.

Figure 6

A summary of the performance achieved when only 8 patients from Egypt, Uganda, Ghana, Algeria and Malawi are used in terms of recall and accuracy. First, the model is directly trained with the African samples (yellow). Second, the model is trained using only the large Spanish dataset (red). Finally, the pre-trained Spanish model is fine-tuned with the new African samples (green).