Radiology: Where AI makes up for manpower shortage

India has been facing a shortage of radiologists for a while now. The deepening of artificial intelligence (AI) capability may help address the problem.

In a recent paper titled ‘Autonomous AI for multi-pathology detection in chest X-rays: A multi-site study in the Indian healthcare system’, which is yet to be peer-reviewed, medical AI firm 5C Network’s CEO, Kalyan Sivasailam, and others have shown that autonomous AI can help halve radiologists’ reporting times.

Why is AI necessary here? “There are only an estimated 10,000-15,000 radiologists for a population of 1.4 billion in India. AI as an aid will help lessen the workload while helping increase accuracy of analyses,” the paper says.

The authors of the paper have outlined the development of an autonomous AI system for chest X-ray interpretation, “trained on a dataset of over 5 million X-rays sourced from healthcare systems across India”. Sivasailam points out that this AI system integrates advanced architectures, including vision transformers, faster R-CNN, and various U-Net models (such as attention U-Net, U-Net++, and dense U-Net) to enable classification, detection, and segmentation of 75 distinct pathologies, ranging from pneumonia to complex conditions.

The tool can help classify X-ray results as ‘normal’ or ‘abnormal’ with high accuracy — 99.8 per cent, 99.6 per cent, and 99.9 per cent, respectively, on the counts of precision, recall, and negative predictive value. ‘Precision’ here means that, of the X-rays the AI identified as normal, 99.8 per cent were verified to be normal, Sivasailam explains.

‘Recall’ implies that the tool correctly identified 99.6 per cent of normal chest X-rays. ‘Negative predictive value’ implies that if the AI tool assesses an X-ray result as normal, then there is a 99.9 per cent chance of it being normal.

The tool also identified a specific pathology correctly up to 98 per cent of the time, and can correctly identify over 95 per cent of the instances of a specific pathology.

Focus on abnormalities

With the AI tool, the team could also detect the location of abnormalities by generating bounding boxes around potential pathology regions. The capability to generate high-precision, pixel-level segmentation of these regions helps in visualising the extent and shape of the pathology.

“By automating the initial analysis, and flagging potential issues, it allows radiologists to focus on more complex cases,” says Sivasailam.

He says that the high precision and recall rates suggest that AI contributes to “improved diagnostic accuracy with its reliability, potentially reducing the number of missed diagnoses”.

But given that the results are not 100 per cent, wouldn’t every report need to be reviewed by a radiologist, bringing into question the usefulness of AI? “Radiologists will save time; they have to just validate what the AI has thrown up. Let’s say a radiologist takes two minutes to read an X-ray; with AI’s help, the time saved to read and analyse the X-ray could be 50-60 per cent.

The radiologist can also afford to spend more time on ‘abnormal’ cases.

We have heard of AI and generative AI. But what is autonomous AI? Explains 5C Network CEO Kalyan Sivasailam, “Autonomous AI refers to systems that can independently perform tasks by integrating multiple AI capabilities, such as perception and decision-making, with minimal human intervention. For example, in medical imaging, computer vision enables a machine to analyse an image and detect abnormalities. Then, generative AI creates a detailed report describing the findings. When these capabilities — perception, analysis, and report generation — are combined, we achieve autonomous AI.”

Elaborating, he says, “Let’s say there is a machine, a black box into which an X-ray image is fed as input. The box first detects the pathology in the image. If it spots any abnormality in the image, that becomes an output for one part of the machine. This is fed as input to another part of the machine, namely the language component, whose output will be a structured report that is generated without a human. This is passed on to a radiologist for validation.”

During deployment, the system processed more than 1.5 lakh scans, averaging 2,000 chest X-rays daily.

Bridging skill gap

In a country like India, where the limited number of radiologists are concentrated in urban centres, leaving rural areas under-served, an AI tool can help bridge the gap.

As for the limitations of the technology, the paper points out that results may not be accurate, for instance, in images with overlapping structures or low-contrast areas, such as dense lung regions. Human radiologists, with their contextual understanding and ability to integrate subtle visual cues, may be better placed to review such reports.

Chest X-rays often need adjustments, such as rotation correction to fix their orientation or contrast adjustments to make details stand out. Without these steps, AI tools designed to analyse the images can sometimes produce inaccurate results. Trained humans, he says, can often interpret images with a wider range of quality and orientations.

Sivasailam points out that this need for image preprocessing before analysis is a key hurdle to be overcome before autonomous AI can become ubiquitous in medical imaging.