AI ML Models Analysing USG Scans for Clinical Decision Support in Perinatal Care

This article was originally published on avijitguha.medium.com

Perinatal, as per the Oxford Advanced Learner's Dictionary, means at or around the time of birth. The Farlex Partner Medical Dictionary defined it as, “occurring during, or pertaining to, the periods before, during, or after the time of birth; that is, before delivery from the 22nd week of gestation through the first 28 days after delivery.” The care provided to a pregnant woman and her baby, after the embryonic stage, during pregnancy, childbirth, and a few weeks after the delivery is a crucial factor towards building a better and safer world. Ultrasound Scans or Ultrasonography (USG) are used to scan images of the foetus, placenta, uterus and surrounding structures in the abdominal cavity. Analysis of a USG scan, like many other medical scans, requires the expert eyes of a radiologist or gynaecologist to decipher various symptoms beyond the overall well-being parameters mapping the gestation period.

?

Artificial Intelligence (AI) components like Computer Vision (CV) and Machine Learning (ML) models, offer new dimensions in analysing data from ultrasound images, thus promising effective and efficient diagnosis and treatment, if required, during the perinatal period. These technologies also provide valuable insights about various healthcare parameters ensuring better health of the mothers and the babies.

Related Studies and Experimental Models

Using machine learning models for the analysis of USG images of foetuses in the third trimester of pregnancy we can predict the risk of preterm birth and various other complexities including congenital conditions. Such models are of great importance for Perinatal Care as preterm birth is a leading cause of maternal and infant mortality as well as morbidity. As we are able to predict the risk of preterm birth, a lot of preventative measures can be planned well in advance; often resulting in saving lives and a healthy mother and child.

Studies show promising results of AI/ML models for various perinatal care applications. For example, we can identify probable defects in foetal development well in advance; even predict the risk of foetal distress by analysing USG data.

One example of an AI/ML model for analysing USG scans is DeepGestalt, developed by researchers at Stanford University. Using this model of ML algorithms one can analyse foetal USG scans and predict probabilities of various abnormalities, such as congenital abnormalities or neural tube defects.

Another example is the AI/ML model developed by researchers at the University of Oxford and the University of Verona uses deep learning algorithms to analyse foetal USG images to predict the risk of preterm birth. This model was able to predict preterm birth with an accuracy rate close to 85%.

AI/ML models for analysing USG scans can also be used to improve the accuracy of foetal growth assessment. A study published in the journal Ultrasound in Obstetrics and Gynaecology found that an AI/ML model was able to accurately predict foetal weight up to an accuracy of 95%. Accuracy in foetal weight prediction can lead to corrective interventions like modifying nutrition with supplementation or other appropriate interventions for ensuring the health of the mother and child. Even, preparing the mother and the healthcare workers for probable complexities around birth. For example, an overweight baby would pose higher risks of perineum ruptures and associated postpartum haemorrhage.

?

AI/ML models can also provide useful insights into the health of the mother and her unborn baby by analysing USG images.?Researchers at the University of California, San Francisco used an AI/ML model to analyse USG scans and predict the risk of placental dysfunction, which could be correlated to complications like preterm birth or foetal distress. This model proved to predict placental dysfunction with 82% accuracy.

Risk of using Machine Learning for Ultrasound Analysis

While AI/ML models for the analysis of perinatal scans promise significant benefits to clinical decision support, there are several risks and ethical issues associated with using machine learning models in medical issues like these. One of the primary risks associated with using ML models in healthcare is the potential for biased or inaccurate predictions. Analysis of USG reports is no exception. It’s a life-and-death situation literally.

Unsupervised ML models used to detect anomalies or identify clusters learning from large datasets to identify patterns or trends in the analysis of USG scans without the use of labelled examples are more prone to false positives or false negatives. For example, an unsupervised model may identify a cluster of suspicious pixels on a USG scan, but this cluster may not actually represent carcinoma. Or, it may fail to identify multiple growths around the ovary ignoring them as fimbriae tubae, leading to a false negative result.

While supervised ML models are better posed with training on labelled examples, there are significant risks when the training data is not representative of the broader population, or when the labelling process itself is biased or ill-conceived. Moreover, supervised ML models may be vulnerable to adversarial attacks. These attacks involve intentionally manipulating the input data in a way that causes the model to make incorrect predictions in favour of the commercial interests of a healthcare facility.

Both unsupervised and supervised ML models have their strengths and weaknesses. However, depending on the specific goals, availability of data in statistically substantial volume and considerations around the associated risks, methodologies like supervised, self-supervised, unsupervised etc. may be deployed keeping the greater cause in mind.

Conclusion

As we can see, the AI/ML models, supervised, self-supervised or unsupervised, for analysis of USG Scans can largely improve the outcomes and efficiency of perinatal care; by detecting difficult-to-diagnose abnormalities and speeding up the process of the diagnosis. The various early alerts possible to generate using these models can be life-saving for both, the mother and the child. Even while there are several risks associated with their use, including reduced accuracy, false positives/negatives, bias, and vulnerability to adversarial attacks.

It is the call of the healthcare providers to consider or construct unbiased ML models for analysing USG images and to ensure responsible and ethical engagement of these models carefully considering probable risks involved in deploying such technologies. Overall, these models hold great promise of superior patient outcomes during the perinatal period.