Latest:
Neeraj Lalwani, MD; Anastasia Fotis, BS; Pankaj Gupta, MD; Judy Yee, MD

Artificial Intelligence in Abdominal Radiology: Transforming Diagnosis, Prediction, and Workflow Optimization

By Neeraj Lalwani, MD; Anastasia Fotis, BS; Pankaj Gupta, MD; Judy Yee, MD

Artificial intelligence (AI) is no longer a distant dream in healthcare it’s here, now, transforming the way we diagnose and treat disease. Nowhere is this more evident than in medical imaging, where AI’s potential is being explored at an extraordinary pace. But beyond hype, what does AI do in clinical practice? And more importantly, is it working?

We recently completed a comprehensive scoping review that mapped out how AI is being used in abdominal radiology, specifically the imaging of vital organs such as the liver, pancreas, and kidneys. Our findings, drawn from 432 studies published between 2019 and 2024, paint a picture of both tremendous progress and lingering gaps, a landscape full of promise but still far from seamless integration.

The Rise of AI in the Imaging Room

AI tools in abdominal radiology can be categorized into four main areas: segmentation, classification, prediction, and workflow optimization.

  • Segmentation refers to the automated outlining of organs or tumors in imaging scans, a tedious task typically done by radiologists.
  • Classification includes identifying whether an image shows signs of disease.
  • Prediction means using imaging to forecast disease progression or treatment response.
  • Workflow optimization covers tools that speed up image processing or help prioritize urgent cases.

Among these, segmentation tools are leading the way, with deep learning models especially U-Net and its variants, showing impressive results in liver and pancreas imaging. Dice scores (a common accuracy metric) ranged from 0.65 to 0.90. These tools are becoming increasingly reliable, particularly in high-contrast scans.

However, performance drops when things get complicated motion blur, low image quality, or overlapping structures can throw off even the best models. Most studies rely on curated, clean datasets. Real-world conditions? Not so much.

AI won’t replace radiologists. But it will change how they work, making diagnosis faster, more consistent, and potentially more personalized.

Diagnosing Disease: How Smart is AI?

When it comes to tumor detection, AI shows real potential. Convolutional neural networks (CNNs) like ResNet and DenseNet can spot abnormalities with up to 100% sensitivity in some studies. But there’s a catch: models perform best on common cancers. For rarer diseases or those with subtle signs, think Crohn’s or autoimmune conditions accuracy takes a nosedive.

A bigger challenge? AI doesn’t think like a doctor. Radiologists interpret images with patient history, symptoms, and lab tests in mind. AI, for now, often works in a vacuum. For real diagnostic power, we’ll need multi-source AI systems that combine imaging with clinical data a technical and ethical challenge that researchers are just beginning to tackle.

Can AI Predict the Future?

The predictive side of AI is perhaps the most exciting and the most underdeveloped. Imagine a tool that could predict whether a liver tumor will grow, or whether a patient is at high risk for an aneurysm rupture. AI models using techniques like XGBoost and LSTM (long short-term memory networks) are showing promise here, with AUC scores (another performance metric) ranging from 0.62 to 0.99.

But clinicians remain skeptical and rightly so. Predictive models are often “black boxes,” offering no rationale behind their forecasts. Without explainability, trust is hard to build. Moving forward, we need explainable AI (XAI) that can communicate its reasoning in ways that clinicians can interpret and validate.

Saving Time, Reducing Repeats — Workflow Wins?

The least sexy but most practical use of AI in radiology is workflow optimization. Here, the gains are clear: studies report 20–30% reductions in report turnaround time and 15–25% fewer repeat scans due to better image quality. AI-assisted triage is also speeding up urgent case reviews by up to 40%.

Yet despite these benefits, most of these improvements aren’t measured using standardized benchmarks. Until we define what success looks like in time saved, errors avoided, or dollars saved workflow AI will remain more of a promise than a guarantee.

So, What’s Holding AI Back?

In a word: validation. More than 80% of the studies we reviewed relied solely on internal testing — AI models were trained and tested on data from the same institution. That’s a recipe for overfitting, where an AI performs well in lab conditions but flops in the wild.

AI must function across many institutions, devices, and patient populations in order to be therapeutically beneficial. That means multi-center studies and external validation expensive, logistically tough, but necessary.

Also missing is research on underused imaging types like ultrasound and X-ray, which are cheaper and more common globally. Most AI tools today focus on CT scans, reflecting where the best datasets exist — but this leaves a huge portion of the world’s diagnostic imaging untouched by AI.

Cost and Ethics: The Other Barriers

AI doesn’t come cheap. Even the most efficient model is useless if it can’t be integrated into hospital IT systems or if it costs more to run than it saves. Smaller clinics may simply not be able to afford it. Plus, the legal and ethical questions around AI, like who’s responsible for a wrong diagnosis, still need clear answers.

The path forward likely involves “human-in-the-loop” systems, where AI assists but does not replace radiologists. These tools can flag abnormalities, speed up routine tasks, and prioritize urgent cases freeing up human experts for more complex decision-making.

Final Thoughts: Not a Revolution, but a Smart Evolution

AI won’t replace radiologists. But it will change how they work, making diagnosis faster, more consistent, and potentially more personalized. Our review suggests that while AI’s promise in abdominal radiology is real, its practice still needs work: better validation, more diverse data, ethical clarity, and a sharper focus on real-world implementation.

As the field matures, radiologists, developers, and regulators must collaborate, not just to build smarter tools, but to make sure they’re trustworthy, accessible, and impactful. The AI wave has arrived. We are responsible for guiding it in the proper direction.

Author’s Information:
Neeraj Lalwani, MD, Professor and Chief of Abdominal Radiology, Montefiore-Einstein, Bronx, NY, USA
Anastasia Fotis, BS, Medical Student, Albert Einstein College of Medicine, Bronx, NY, USA
Pankaj Gupta, MD, Additional Professor, Postgraduate Institute of Medical Education and Research, Chandigarh, India
Judy Yee, MD, Professor and Chair of Radiology, Montefiore-Einstein, Bronx, NY, USA